feat(models): add Phi-4 Multimodal text-only adapter

[zitai-wang]

Summary

• Add a text-only adapter for Microsoft Phi-4-Multimodal-Instruct using AReno native layers and runtime.
• Implement Phi partial LongRoPE, true tied input/output embeddings, text generation, and text-only save/reload.
• Load the official base-language checkpoint with explicit vision, speech, and LoRA classification.
• Add TP-aware packed-section mapping for fused QKV and gate/up projections, validated with TP1, TP2, and TP4.

Validation

• Phi-4 focused CPU tests: 39 passed.
• Full repository suite with CUDA hidden: 643 passed, 15 skipped, 1 failed.
  • The sole failure is tests/test_mlx_training_cpu.py::test_mlx_decode_progress_matches_cuda_log_shape.
  • This MLX logging failure reproduces unchanged on clean upstream/main at 4cc3d388d17b60b534c57d25f69fdcad393433dc; it is not introduced by this PR.
• Available CUDA tests on an NVIDIA GeForce RTX 3090: 12 passed.
• Ruff lint/format, all manual pre-commit hooks, compileall, and git diff --check: passed.
• Official BF16 checkpoint audit: 194 base-language keys consumed, 0 unknown keys, and 0 missing required base-language keys.
• HF/AReno parity: exact 16-token greedy and KV-cache token parity; last-token logits cosine similarity 0.999983340331.
• Tensor parallel parity: exact greedy token parity for TP1, TP2, and TP4.
• Text-only save/reload: bitwise-equal logits and true weight tying preserved.

Known limitations

• This PR supports only the text path. Image and audio support will be submitted in follow-up PRs.
• Cached generation crossing the Phi LongRoPE 4096-token regime boundary currently fails fast by design.