feat(models): add Phi-4 multimodal audio support

[zitai-wang]

Summary

• add native Phi-4-Multimodal audio processing and placeholder expansion
• implement the official causal Conformer audio encoder, Nemo Conv subsampling, T5 relative attention bias, and speech/vision projector branches
• support Speech LoRA activation and independent recurrent state across chunked prefill and KV-cache decode
• load and save the complete official audio tower, projector, buffers, and TP-aware Speech LoRA weights
• preserve PR1 text and PR2 vision behavior, including mixed text/image/audio batches

This is a stacked PR based on feat/phi4-vision-adapter (PR2). After PR2 is merged, this branch should be rebased and the PR target adjusted to upstream/main.

Validation

• official Phi-4-Multimodal checkpoint: 2047/2047 tensors consumed; missing=0, unknown=0, skipped=0
• audio projector and merged embeddings: exact HF parity (max_abs=0)
• last-token logits: cosine 0.9999907, matching top-1 and 5/5 top-5 overlap
• greedy KV-cache generation: 8/8 token ids match HF; audio encoder runs once during prefill
• real checkpoint TP1/TP2/TP4: matching 8-token greedy generation
• multi-audio: different-length segments produce 44 + 22 matching placeholder slots
• full save/reload: 2047/2047 tensors consumed, logits max_abs=0, matching generation
• Phi/multimodal focused tests: 82 passed
• repository suite excluding three confirmed upstream baseline failures: 686 passed, 3 skipped
• available CUDA tests: 12 passed
• pre-commit, Ruff lint/format, compileall, pip check, and git diff check passed

The three excluded repository failures reproduce on clean upstream/main@4cc3d38: one MLX log-capture assertion and two Gloo/NCCL tests that select duplicate CUDA devices in this environment.

Known limitations

• inference support only; training and rollout reward/logprob integration remain out of scope
• cached generation crossing the inherited Phi LongRoPE 4096-token boundary continues to fail fast