if you want to disassemble a conventional deep groove bearing, the load required to displace the centre race off centre (to allow the last ball to be fitted/removed) isn't as high as you might think because the balls make a shallow wedge angle that (relatively easily) flexes the outer race.
For example if the balls occupy an arc of 185 degrees, this only requires that (for a bearing that has nominally zero clearance, which BTW many do not) the diameter of the raceway increases by ~0.1% where the first and last balls bear against it.
FWIW for a lot of bicycle applications, it would be better if the bearings were constructed as "full complement" types, which have almost twice as many balls, and are assembled via a loading slot/special raceway. There are several ways of assembling such bearing:
a) One of the two raceways has one shoulder almost completely removed; this allows the groove in the intact raceway to be filled with balls, and then the raceway with the truncated shoulder to be fitted. Often there is enough shoulder to make the assembly process require a load, and/or there is a snap ring in a groove that keeps the bearing together. The resulting bearing then can only be used as one half of a pair on a shaft, each of which is used as an angular contact bearing. Bearings in Surly hubs are angular contact bearings that are made this way, using a tiny snap ring to prevent the bearing from coming apart.
b) A simple loading slot is used, in which each raceway has the shoulder partly removed so that the balls can be fitted. This may or may not require force to fit the final balls, depending on the geometry of the loading slot. Again a snap ring may prevent the bearing from coming apart. Full complement bearings that are (optionally) used in suspension pivots on FS MTBs are made this way. In near-static (or at least limited articulation) applications like that, you can install the bearing so that the loading slots never line up with one another; this prevents the balls from escaping, even should the bearing become worn.
c) A third type (for a directional bearing) is a variant on the loading slot described above, in which the slots are angled, such that when the bearing is rotated forwards the balls roll past the loading slots at an angle (a bit like a train running through a set of points), and therefore cannot escape through them, unless the bearing is both worn and is turned backwards. Bearings of this type can be used through a full range of angles (i.e. rotated continuously) even if there is a small amount of free play.
All these bearings have a higher load capacity and a lower speed rating than a standard deep groove (Conrad) bearing, in which the balls are separated by a clip of some kind. The higher load rating is achieved simply by sharing the service load between more balls. The lower speed rating arises because the balls rub against one another (which is apparently worse than rubbing against a clip at high speed). Types b) and c) also allow the balls rolling past the loading slots to be partly unloaded and this can cause more rubbing between balls and slightly more wear to arise.
Standard Conrad bearings (especially those which are operated at high speed) are intended to have a preload on them at all times, so that the balls remain evenly spaced and don't rub hard against the clip all the time. If this preload is excessive the bearing life is reduced (due to simple overloading) and if it is lost the bearing life is also shortened because the balls rub against the clip and it is only a matter of time before the bearing fails completely.
Most deep groove (Conrad) bearings that are of a size that might go into hubs, pedals, BBs etc have speed ratings of several thousand rpm; by contrast in a bicycle application they are unlikely to be doing more than a few tens or low hundreds of rpm. This means that the operating conditions are almost never in the regime in which a deep groove ball bearing is designed to operate, in which the metal parts are separated by a dynamic lubricant film and don't actually touch once another. In fact in bicycles, cartridge bearings are often turning up to x1/100th the maximum speed they were designed for; this means that using different lubricants, different seals, and different bearing designs (eg full complement types) is not only fair game, it is a really good idea.
Arguably a well-made cup and cone bearing is an idealised type of full complement angular contact bearing. Better than most industrial bearings, variations in fit and finish of the parts can be accommodated because they can be adjusted. Furthermore they can be adjusted to whatever degree of precision you can be bothered to apply to them.
Note that in any full complement bearing, you can either set up the bearing with or without preload. With preload (that is comparable to the service loads), if the balls make contact with one another, they will do so under load and this promotes wear and drag in the bearing. However if there is no preload, the balls are only in contact with one another when one or both is actually loose, i.e. unloaded; this is, at low speeds, relatively benign in terms of wear. The balls can only scuff against one another or against the raceways under load if they are not positioned precisely as they come into that part of the bearing in which they are loaded; this in turn can only happen if the bearing is adjusted so that there is too much slack.
As an experiment, you can try assembling an unsealed cup and cone bearing front hub with just a little light oil in the bearings. What you will probably find is that as you ride, you can hear the balls 'rolling over the top' of the cone (where they are unloaded) and falling down and hitting the other balls, making a kind of 'ticking' sound. They are only in loaded contact under the bottom of the cone, and the unloaded contact does not cause appreciable wear at these low speeds with lubricant present. If preload is applied, you won't hear the same 'ticking' noise' the balls are in hard contact all the time; (silent -at first- but ultimately deadly). If the bearing is adjusted with too much slack, you may hear the balls scuffing against the raceway, as they come into the loaded part of the bearing; this can make a noise that is difficult to describe but is not unlike the one you hear when walking on a shingle beach. The difference between these three conditions ('just right', 'too much preload', and 'too much slack') is potentially as little as 10um each way, or plus or minus about three degrees of cone adjustment. Note that a typical axle compresses by about 80um when the QR is used to tighten the wheel in the frame; if you don't allow for that the bearings are guaranteed to be overly preloaded.