result_caching/copying looses StimulusSet attributes that are not in `pd.DataFrame._metadata`

[KogniJannis]

The _place_on_screen function in screen.py adds the paths to the untransformed images after placing images on the screen as a custom StimulusSet attribute:
l.72 converted_stimuli.original_paths = copy.deepcopy(stimulus_set.stimulus_paths)
It seems to me that when result_caching's @store pickles and unpickles the StimulusSet, all attributes not defined in _metadata here in brainscore_core are lost.
Thus, the transformed stimulus set is different when it is computed vs when it is loaded from cache.

This is also the case when the StimulusSet is copied or sliced, see this older StackOverflow entry.

I believe this behavior is completely silent right now, because a) nothing actually uses the original paths, b) there is no test that checks whether result_caching returns identical attributes. Therefore, there is really no urgency to address this right now, but I'd still say it is unexpected/unintended behavior.

Steps to reproduce: load a transformed stimulus set from .result_caching/brainscore_vision.benchmark_helpers.screen._place_on_screen and check for original_paths

Possible ways to address this might be to either

• extend _metadata
• use converted_stimuli.attrs['original_paths'] (it seems to me that then the attribute will survive pickling, but it is not a proper attribute so any hasattr check will still fail (?))
• or address it over at result_caching by making a custom store decorator for DataFrames/StimulusSets that preserves any attribute