[AndersenAgnostic] Fix RegionAwareMemoryNodeProvider, now based on call graph and handling calls to external

[haved]

Introduces a call graph creation step to the RegionAwareMemoryProvider, which ensures that all possible call paths are handled, even when going via external function calls.

SCCs are found in this call graph, and they are visited in reverse topological order.
This means that for any function call found during the "annotation" step, it either points to a function in an earlier SCC which is already finished, or in the same SCC. In the latter case a flag is used to remember that the call may be recursive.

RegionSummaries also remember if they may contain recursion. Once all functions have been annotated, a final pass goes through all summaries that may contain recursion, and add the whole SCC's summary of memory locations to it.

These changes made the encoding more precise in some of the tests, so they have been updated to account for that.

Testing the llvm-test-suite with --AAAndersenRegionAware passes on all tests, though I managed to get the encoder to fail when combining --AAAndersenRegionAware with other passes. I felt this PR was large enough already, so fixing encoder issues is not done here.

I am adding the [AndersenAgnostic] on this PR, but that is only because [AndersenRegionAware] does not exist yet.

[github-actions]

The execution time of dynamatic/pivot has changed
  Golden cycle time: 1190
  Simulated cycles: 1181

[phate]

@haved I am excited to review it, but is there any chance that we could move the iteration wrapper changes into its own PR?

[haved]

@phate that is absolutely possible, I can do that

[github-actions]

The execution time of dynamatic/pivot has changed
  Golden cycle time: 1190
  Simulated cycles: 1181

[github-actions]

The execution time of dynamatic/pivot has changed
  Golden cycle time: 1190
  Simulated cycles: 1181

[phate]

The first batch of comments. I need to go over it one more time, as I was not able to finish it completely.

[haved]

@phate I addressed your comments so far

[github-actions]

The execution time of dynamatic/pivot has changed
  Golden cycle time: 1190
  Simulated cycles: 1181

[phate]

Great work! The code is very readable and understandable.

There is still the empty TestIteratorWrapper.cpp in the PR that needs to be removed.

I have a question: I do not see that we handle setjmp/longjmp explicitly. Do we handle it correctly?

[haved]

@phate I created a Context to make it easier to clear temporary state (and actually clear it).

Regarding setjmp and longjmp, they are currently handled, as we are currently very conservative within each SCC, and anything that calls either will end up in the big "external + almost everything else"-SCC.

[github-actions]

The execution time of dynamatic/pivot has changed
  Golden cycle time: 1190
  Simulated cycles: 1181