Stop unloading the profiler on shutdown to prevent segfaults on background threads

[davmason]

Fixes #26687

On Linux we were seeing profilers crash in ELT hooks because the profiler would be shut down before all managed threads had exited. These checks will prevent slow path ELT from crashing, profiler vendors will have to deal with fast path ELT calls that come in after the profiler is shut down.

[jkotas]

What happens if the profiler is torn down right after we do this check, but before we actually call it below?

[davmason]

Yeah, there are all sorts of races here. I realize now that even the check can race, depending on the order of events g_profControlBlock.pProfInterface could be set to null after g_profControlBlock.pProfInterface.Load() == NULL and then g_profControlBlock.pProfInterface->GetEnterHook() == NULL would AV. I need to rework this entirely.

[davmason]

Looking at why this happens on Linux and not Windows it's because we don't set g_fProcessDetach on Linux. Talking to @janvorli this is on purpose, and we should try to avoid doing any checks about why the runtime is terminating. We already would skip freeing the profiler in many cases on Windows, I think the only safe thing to do here is to stop freeing the profiler on shutdown altogether. It may affect some profilers, but I believe they will be able to either free what they need to in the Shutdown callback or when the OS frees the library after we exit.

The only other options are to continue crashing sometimes or to put expensive synchronization on a hot path in ProfilerEnter,

[noahfalk]

LGTM, sorry I missed this earlier