[jig-mistry]

Closes #1167.

After a model has been called with require_variance=True and more than one trial, the next call without require_variance fails:

  File ".../brainscore_vision/model_helpers/activations/core.py", line 413, in _package_layer
    assert layer_activations.shape[0] == len(stimuli_paths) * runs_per_image
AssertionError

__call__ only updates _microsaccade_helper.number_of_trials on require_variance=True calls, so the count from the last such call sticks around. _get_activations_batched then runs every stimulus that many times, even with require_variance off. But _package_layer expects one run per stimulus.

Two ways to get there:

• the one from the issue: a benchmark that fits with require_variance=True and then calls look_at without it;
• one loaded model run through Malania2007 and then another benchmark.

Existing scores aren't affected. The bug raises rather than returning wrong activations, and scoring rarely reaches it:

• No benchmark does the first today. Malania2007 is the only one that uses require_variance, and it uses it for both fitting and testing.
• brainscore_vision.score calls load_model for each benchmark, and nearly every plugin builds a new extractor there. (resnet50-moclr8deg is an exception: it builds its wrapper at import and reuses it.)

The fix: _get_activations_batched now only repeats stimuli when require_variance is set, as _package_layer and attach_stimulus_set_meta already do. This covers __call__ and direct from_paths / from_stimulus_set calls alike. Calls with require_variance=True that work on master give the same results.

I went with this rather than resetting the count in __call__, as the issue suggests, for two reasons:

• A reset would miss direct from_paths / from_stimulus_set calls.
• LayerPCA fits itself by calling the model again from inside a batch hook (pca.py:47). If that happens during a require_variance=True call, the nested call fails on master too, because it inherits the outer trial count. This change fixes that case. A reset would only move the failure to the outer call, since the nested call would reset the count under it.

Test: test_number_of_trials_does_not_carry_over makes a require_variance=True call, then calls the same model without it through __call__, from_paths and from_stimulus_set. It fails on master with the error above and passes here.

Related: the stored-activations cache key always records number_of_trials=1, whatever the actual trial count. That's in a separate PR, #2629, since it changes the cache keys of require_variance=True requests.