fix: accept PyTorch tokenizer inputs in MLX generate

[Vaishnavi220506]

Problem

On macOS, AutoModel.from_pretrained() routes models through AirLLMLlamaMlx. generate() passed the PyTorch tensor returned by the standard tokenizer directly to mlx.nn.Embedding, which rejects it before generation starts.

Changes

• Normalize input IDs at the generate() boundary.
• Preserve existing MLX arrays without copying them.
• Convert PyTorch tensors through a detached CPU NumPy view before creating the MLX array.
• Add regression coverage for PyTorch tensors, MLX arrays, and Python sequences.

Validation

• python -m compileall -q air_llm/airllm/airllm_llama_mlx.py air_llm/tests/test_mlx_input_types.py
• Targeted repository tests: 47 passed, 3 MLX tests skipped on Windows.
• python -m build
• python -m twine check dist/*

Closes #280