Fix compiler crash in HeapToStack optimization pass

[SeanTAllen]

This is a good one. One the March 22, 2022 development sync,
we collectively debugged issue #4059. We finally arrived at
the conclusion that our call graph usage was incorrect.

Originally during the call, Joe realized that while we were
using, we weren't doing it correctly. The initial assumption
was that the incorrect usage was innocous, however, it was
determined shortly thereafter that something about the incorrect
usage was in fact causing issue 4059.

This commit removes the offending code.

Gordon reported during the call that when the call graph code
was added as part the (now years ago) addition of LLVM 4 support
that the code was cribbed from the Dlang compiler and he posited,
probably not correctly.

Fixes #4059

[jemc]

For posterity, I'll summarize my key takeaway from the call yesterday:

If we wanted to bring back this call graph analysis maintenance code, we'd want to first fix whatever the unknown bug in it is.

Then once it was working, in order to actually get the benefit of maintaining the call graph analysis, we'd want to tell LLVM that we maintained it by marking it as a "preserved analysis" in the analysisUsage function. Otherwise, LLVM is just going to discard the analysis and regenerate it anyway.