linker/inline: handle OpPhis.

[eddyb]

This is a prerequisite for:

• Minimally integrate SPIR-T (opt-in via RUSTGPU_CODEGEN_ARGS=--spirt). #940 (comment)
  ^{(the link goes to my comment explanation the plan that this PR is a part of)}

The problem this PR solves is that the inliner did not support "φ nodes" (OpPhi in SPIR-V), as they did not exist at the point where it was ran (i.e. all passes that would introduce OpPhis ran after the inliner).
This manifested as broken OpPhis, since the sources of control-flow edges (i.e. the side with the branch instruction) changed without the OpPhis in the destinations being updated to reflect the CFG change.

But with SPIR-T, we want the speedups that come from being able run the inliner after passes that create OpPhis (as the SPIR-T CFG structurizer can handle OpPhis just fine, unlike our existing one).
More specifically, in #940, we end up with two mem2reg passes, trying to do as much as possible before, and the rest after, inlining (though ideally they it should be interleaved with inlining, as per EmbarkStudios/spirt#6).

Thankfully most of the changes (in the big inline_block function) were only for readability, and the bulk of the additions just change some basic-block IDs in place, they don't do any complicated analysis or rewriting, so it felt straight-forward enough to have it in the existing inliner, instead of trying to replace it with a SPIR-T one.