GH-150516: Reduce the work done to spill and reload the stack around calls

[markshannon]

Avoids needless spilling of the stack around calls, as described in the issue.

Also adds the stackpointer_valid to frames in the debug build to verify that the frame's stack pointer is not used when it shouldn't be. frame->stackpointer_valid is equivalent to the old frame->stackpointer != NULL check.

The new scheme is more precise, so I needed to change a few stack saving/syncing/reloading operations to better reflect the true semantics.

• Issue: Streamline spilling around calls in the code generator #150516

[markshannon]

Skipping news as this has no impact on any users

[eendebakpt]

Here is a latent issue:

inst(TEST_LOOP_SPILL, (a -- a)) {
    int n = (int)oparg;
    for (int i = 0; i < n; i++) {
        escaping_inside_loop(a);
    }
    escaping_after_loop(a);
}

The generator ellides the _PyFrame_SetStackPointer for the escaping_after_loop, which can be a problem if n is zero.

[markshannon]

That's a pre-existing issue where we don't merge stacks properly at the end of for loops.
It needs fixing though, so I've added a fix for that here.

[eendebakpt]

For proper benchmarking of this PR a PGO build is needed. I did a subset of 19 loop-heavy benchmarks from pyperformance and got +1.9% geomean (biggest win is deltablue at 1.10x).

[markshannon]

Our numbers for the whole pyperformance suite show a speedup, but less than 1%.

[markshannon]

The fuzzers are producing compile errors that make no sense

[diegorusso]

This has been reviewed internally and LGTM.