Feature/bart adapter

[hsusul]

Description

Adds initial TransformerBridge support for BartForConditionalGeneration.

This implements a minimal BART encoder-decoder adapter following the existing T5/MT5 encoder-decoder patterns, with BART-specific module paths for shared embeddings, encoder blocks, decoder blocks, decoder cross-attention, fc1 / fc2 MLPs, layer norms, and lm_head.

Fixes #1401

This first slice is intentionally raw-HF parity focused:

• supports BartForConditionalGeneration
• supports symmetric encoder/decoder BART configs
• maps decoder cross-attention with is_cross_attention=True
• disables fold-LN / center-writing processing for post-LN BART
• does not add Pegasus, mBART, Marian, or M2M100 support
• does not implement compatibility-mode weight processing

No new runtime dependencies are required.

Type of change

• New feature (non-breaking change which adds functionality)

Screenshots

N/A

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

[jlarson4]

Great work on this! Merging