Remove BatchT/OutputT generics, introduce ReconstructionLoss protocol

[ocg-goodfire]

Description

This PR was authored by Claude (Opus 4.6) at Oli's request, implementing the de-genericize plan from PR #363.

Removes BatchT/OutputT type parameters from ComponentModel and all downstream code. Model batch input and output types are now Any. Introduces a ReconstructionLoss protocol with concrete implementations (recon_loss_mse, recon_loss_kl) that callers pass explicitly, replacing the old output_loss_type string config.

Key changes:

• Remove BatchT/OutputT generics from ComponentModel and all downstream code
• Add ReconstructionLoss protocol in spd/models/batch_and_loss_fns.py with recon_loss_mse and recon_loss_kl
• Rename pretrained_model_output_attr → extract_tensor_output with regex-based accessor parsing (supports .logits, [0], .output[0], etc.)
• Remove output_loss_type config field (added to deprecated keys with migration)
• Remove extract_batch_data utility — callers handle tuple extraction inline
• Add lm_collate_fn for LM data loading in spd/data.py
• Update all 24 YAML configs, 11 metric files, 4 experiment scripts, and all tests

Related Issue

Implements the de-genericize portion of #363

Motivation and Context

The BatchT/OutputT generics on ComponentModel added type complexity without meaningful safety — model I/O is inherently heterogeneous across experiment types. The output_loss_type string-based dispatch was a code smell; passing a concrete ReconstructionLoss callable is simpler and more extensible. The extract_tensor_output rename + regex parser is more general than the old pretrained_model_output_attr approach.

How Has This Been Tested?

• make check passes (basedpyright + ruff lint + ruff format, 0 errors)
• make test passes (359 passed, 16 skipped, 0 failures)

Does this PR introduce a breaking change?

Yes — configs with output_loss_type or pretrained_model_output_attr will use the migration path in configs.py (deprecated key handling + value conversion). The output_loss_type field is removed from Config; callers must pass reconstruction_loss explicitly to optimize(). The extract_batch_data utility is removed.