feat: Sparse probing eval based on "Are SAEs useful" paper

[chanind]

This PR adds a sparse-probing eval called sparse_probing_sae_probes to keep it separate from the original sparse-probing eval in SAEBench. This eval is based on the SAE-Probes paper Are Sparse Autoencoders Useful? A Case Study in Sparse Probing. The benefit of the sparse-probing tasks from this paper are the following:

• Lots of datasets: The SAE-Probes paper evaluates on over 140 sparse probing datasets
• cross-validated probing: The SAE-Probes paper optimizes the probing pretty heavily to give a stronger realistic baseline to compare against.

This benchmark wraps the standalone sae-probes package, putting results in SAEBench format.