Add optional index group filter to SourceData.one_key()

[philsmt]

During the offline data reduction trials, calls to SourceData._get_index_group_sample() ended up being a significant performance problem, as the current implementation triggers a call to the fairly expensive SourceData.keys().

Initially this was fixed by overwriting this method only within exdf-tools, this MR means to upstream these changes now.

Since the improved implementation is based on SourceData.one_key(), I decided to augment this method by an optional keyword argument index_group and entirely replace the usage of _get_index_group_sample().

Something I noticed in the implementation is that FileAccess.get_one_key() (and in extension SourceData.one_key()) may technically return None if a source has no keys. While the selection mechanism prevents no keys to be selected, it does not seem to prevent files with key-less sources. I would not expect the current DAQ or calibration to create such files, but it's not impossible for someone external to prepare them. For now I decided to not change any behaviour, but preserve the one from _get_index_group_sample() and raise an exception if asking for a non-existing index group.

[philsmt]

I separated the changes to FileAccess and SourceData in separate commits for easier review.

[takluyver]

Other than that one comment, LGTM.

I agree that it makes sense to throw an error for a non-existent index group. :+1:

[philsmt]

I added two additional minor changes for FileAccess:

• Fixed the branch on line 500 to test for None identity rather than boolean False
• Fixed the test to try once after asking for keys (to hit cache) rather than after first ask (which does not populate the cache)

Will merge later today unless something comes up.