Rewrite exp list history to avoid using inspect.stack

[dwpaley]

Calling inspect.stack() appears harmful in a medium-sized (512 ranks) MPI job because every rank hits the filesystem to resolve module absolute paths. This was causing hangs of up to 5 minutes. Here we switch to sys._getframe to work around the problem.

[dagewa]

Thanks @dwpaley. Reading https://docs.python.org/3/library/sys.html#sys._getframe I see that sys._getframe "should be used for internal and specialized purposes only" and that it is not guaranteed to exist. Well, I guess this use does qualify as specialized, but perhaps we should catch if it doesn't exist and fall back to inspect.stack() in such cases.

I want to test this in a few production environments (dev build DIALS, release build DIALS, CCP4 DIALS, conda DIALS...) just to be sure we are getting what we expect. Might take a little while to get on to it, but it's on my list.

[dagewa]

Success with DIALS (cmake) bootstrap build, release build, and DIALS installed from conda. Can't easily check CCP4 DIALS yet as it's on DIALS 3.21 and doesn't have the required code for the experiment history. But I think if there's any problems there we can make a special case for that later on.

So, I'm happy, but I think a fallback to inspect.stack() is a good idea if someone installs DIALS in some weird environment where sys._getframe does not exist. Plus needs a newsfragment.

[ndevenish]

CPython implementation detail: This function should be used for internal and specialized purposes only. It is not guaranteed to exist in all implementations of Python.

This looks like we are fine, because it seems to imply it is always present in CPython. However, we should at the very least check _getframe exists and throw our hands up (or just do nothing with history)