use xarray netcdf storing for all cached results

The scores of models on benchmarks or ceiling scores for benchmarks are cached using https://github.com/brain-score/result_caching's `@store` method. `@store` uses pickle to serialize and deserialize objects for disk IO which has issues with forward compatibility. To keep things better compatible, we should use [store_xarray](https://github.com/brain-score/result_caching/blob/94cb58c8b3a0a993aec9423122b62928af5e1812/result_caching/__init__.py#L465)

Start by locating the benchmark and ceiling-score cache uses of result_caching's @store method. Compare those call sites with the referenced store_xarray implementation, then update the cached-result paths to use it and verify that storing and loading all affected results remains compatible.

Written by the indexing model from the issue text.