Support gradient tracking for MoE models with fused-parameter experts/router

Bergson currently only tracks nn.Linear, HF Conv1D, and nn.Conv{1,2,3}d modules (see LayerAdapter.supported_modules in bergson/gradients.py:238, known limitations section of focs). Modern MoE models with fused-parameter experts/routers (bare nn.Parameters rather than nn.Linear layers — e.g. gpt-oss, Mixtral, Qwen-MoE, OLMoE in transformers 5.x) are silently skipped.

We should add gradient-collection support for these fused MoE layouts so experts and routers are attributed correctly.

Notes:

- LayerAdapter (bergson/gradients.py:237-262) is the extension point.
- Should cover GPT-OSS, Mixtral, Qwen-MoE, and OLMoE.-
- Account for per-token expert routing/sparsity when computing per-example gradients.
- Possibly the simplest solution involves letting the user manually specify bare 2D nn.Parameters, or find the 2D nn.Parameters...

We could also add support for VAEs etc.

Start in bergson/gradients.py at LayerAdapter.supported_modules (around lines 237-262) and review the documented limitations referenced in the issue. Compare the fused-parameter layouts of GPT-OSS, Mixtral, Qwen-MoE, and OLMoE, including per-token routing and sparsity, then define completion as correct expert and router attribution in per-example gradients for those models.

Written by the indexing model from the issue text.