Fix DeepSeek V2 dense MLP hook semantics

[koriyoshi2041]

Closes #1645.

DeepSeek V2 uses dense gated MLPs for the prefix selected by first_k_dense_replace, but the adapter currently gives every layer MoE boundary aliases. This adds optional dense projection bridges and resolves the compatibility aliases to gate_proj, up_proj, and down_proj when those modules exist, while sparse layers keep their existing block-boundary aliases.

The regression coverage checks exact hook identity and width on a tiny dense layer, and verifies that a later sparse layer retains its current behavior.

Tests:

• uv run pytest -q tests/unit/model_bridge/supported_architectures/test_deepseek_v2_adapter.py tests/integration/model_bridge/test_deepseek_v2_adapter.py (44 passed)
• uv run mypy transformer_lens/model_bridge/supported_architectures/deepseek_v2.py
• make check-format
• git diff --check

[jlarson4]

Thanks for taking this on @koriyoshi2041! Looks good to me, just one note about a bug in the test suite.

[jlarson4]

Thanks for fixing that up @koriyoshi2041! Great work, merging now