Support module queries; handle NaN grads; refactor Attributor

[luciaquirke]

Cleanup/small fixes:

• Perform sanity check for NaNs in index grads (shouldn't happen)
• Resolve post unit norm NaNs in Attributor (sometimes happens when the grads have been converted to zero vectors by the final dtype conversion before saving to disk)
• Support queries where k > n
• Support querying module/s in addition to full model grads
• Use a less performant default FAISS config that works on all devices. The recommended one is documented in the docstring.
• Move FAISS logic into its own file to simplify Attributor
• Add FAISS CLI flag to query_index
• Deprecate unstructured gradient indices
• Harden check for existing FAISS index
• Remove redundant unit norm
• Rename chunk to shard

[luciaquirke]

Smaller default dataset for casual consumers/prototyping

[norabelrose]

just use epsilon instead of nan_to_num and then I think it's good to go

[norabelrose]

hmm I guess this works although the standard way to do this is to add epsilon, like 1e-8. It is a hyperparameter but it means the function isn't sharply discontinuous around zero