cfg: reuse original ControlRegions as much as possible during structurization.

[eddyb]

Original motivation back when I first started this refactor was this bug:

• SPIR-T panics with assertion failed: list_eq_key(children) ==\n list_eq_key(self.func_def_body.at(unstructured_region).def().children). rust-gpu#1086

Progress stalled back then and I only revived the branch for the Vulkan layer (#53), because trying to run the Vulkan CTS with the SPIR-T layer would run into:

• missing OpSpecConstantOp support (already implemented in the Vulkan layer branch)
• structurizer region desync list_eq_key panics (fixed by this PR)
• missing OpTypeForwardPointer support (not implemented yet, but it's much rarer and only used with PhysicalStorageBuffer pointers, since logical pointers don't allow recursive data types in the first place)

On top of unblocking the SPIR-T Vulkan layer for in-depth testing, this refactor, out of necessity of not introducing even more spurious unused bools, makes the existing ones as lazy as possible (within the limits of what the structurizer knows without adding more analysis passes).

You can see in the README.md example, that this PR removes both extraneous bools (for "does break" vs "doesn't break"), and instead the original condition (v2) is now used directly by the loop.

I've also theoretically come up with even more advanced techniques (such as taking N disjoint targets that can be dispatched all at once, and lazily generating the appropriate propagation of integer values 0..N to use a switch), but they either require other PRs to land as well first, and/or are too invasive and this has already been a ridiculous sunken cost compared to the measely benefits it offers.

[eddyb]

Just did a once-over, this seems ready to land, just needs CHANGELOG (heh @fornwall beat me to it).

The merge-base changed after approval.