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

[zitai-wang]

Summary

Depends on #521.

• add the native Phi-4-Multimodal SigLIP vision encoder and HD projector
• merge projected image features using the official Phi image placeholder layout
• support Vision LoRA across the language backbone
• preserve per-sequence Vision LoRA state across chunked prefill and KV-cache decode
• support multi-image, variable-resolution, and mixed text/image batches
• load and save the vision encoder, projector, separators, and Vision LoRA weights
• support TP-aware Vision LoRA checkpoint sharding for TP1, TP2, and TP4

Validation

• Phi vision/checkpoint tests: 30 passed
• repository CPU suite: 662 passed, 3 skipped; 1 MLX baseline failure reproduced on clean upstream/main
• CUDA tests: 12 passed
• projected vision features and merged embeddings match the HF reference exactly
• 8-token KV-cache greedy generation matches the HF reference
• TP1/2/4 produce identical greedy tokens; gathered last-token logits remain aligned
• save/reload consumes all 904 supported language and vision tensors
• logits are bitwise identical before and after reload (max_abs=0)
• Ruff, formatting, pre-commit, compileall, and git diff --check pass

Validation used Microsoft Phi-4-Multimodal-Instruct in BF16 on RTX 3090 GPUs. The local snapshot corresponds to ModelScope commit 7641bf905e6965ee54166808d275266371e28343 and Hugging Face commit 93f923e1a7727d1c4f446756212d9d3e8fcc5d81.

The MLX failure is test_mlx_decode_progress_matches_cuda_log_shape: the log is emitted to stderr but is not captured by caplog. The same failure reproduces at clean upstream/main commit 4cc3d38.

Known limitations

• this PR supports text and image paths only; audio support is deferred to a follow-up PR
• cached generation crossing the Phi LongRoPE 4096-token boundary continues to fail fast
• small BF16 hidden-state differences versus HF can accumulate because AReno uses its native attention/runtime implementation