Add forward-only LLaDA architecture adapter

[hsusul]

Summary

Addresses #1465.

Add TransformerBridge support for LLaDAModelLM single-pass inference.

• Map LLaDA residual, attention, MLP, normalization, and logits paths.
• Preserve bidirectional attention and padding-mask behavior.
• Expose standard TransformerLens hook points.
• Require explicit trust_remote_code opt-in.
• Reject unsupported causal generation and shifted-token loss.
• Add offline numerical-equivalence and framework regression tests.

Scope

This PR supports a single masked-token forward pass through TransformerBridge.

Iterative diffusion generation, remasking, confidence schedules, and diffusion-timestep hooks remain out of scope.

Validation

• Offline tiny-model numerical equivalence
• Bidirectional and padding-mask behavior
• Intermediate hook equivalence
• Full state-dict mapping coverage
• FP32 and BF16 coverage
• Existing causal-model regressions
• make unit-test: 3,065 passed

[jlarson4]

@hsusul Great work on this! Looks good, the limitations seem reasonable for initial support. I merged the latest dev into your branch to resolve any conflicts, and will merge once the CI passes