[llvm] [RegisterCoalescer] Deferring deletion of instructions in `ErasedInstrs` until the end of an iteration (PR #79820)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> > > Also can you add the test from #71178 and #73519
> > 
> > 
> > Hmm, this PR doesn't fix it. This test case isn't stable, and I can't test it with RegisterCoalescer.
> 
> I think it's a similar but different issue.

I think it's the same; it's from relying on the delete inst pointers from removePartialRedundancy

https://github.com/llvm/llvm-project/pull/79820