Add Nanbeige (Nanbeige4.2 looped transformer) model support

[MercuriusDream]

Adds the nanbeige model type used by Nanbeige/Nanbeige4.2-3B.

The architecture is a llama-style decoder executed num_loops times with shared weights ("Looped Transformer"). Each loop gets its own KV cache entries and the final RMSNorm is applied at the end of each loop (or once at the end when skip_loop_final_norm is set).

Correctness:

• Logits match the reference modeling_nanbeige.py (fp32, CPU) to a max abs diff of 9.5e-05 over the full 166k vocab.
• Cached token-by-token decoding matches single-shot prefill.
• Verified generation for bf16 and 2/3/4/5/6/8-bit quants on an M-series Mac.

Config flags of the reference implementation not used by the released checkpoint (enable_double_loop_split, loop_share_kv, enable_depth_attention) raise NotImplementedError instead of loading silently. Checkpoints using n-gram embeddings or mHC fail naturally on unexpected weight keys.

Added a test_nanbeige case to tests/test_models.py; the suite passes apart from the pre-existing test_ssm failure on main.

[MercuriusDream]

Is anyone there?

[zcbenz]

Would mind rewriting the PR description following our new template?
https://github.com/ml-explore/mlx-lm/blob/main/.github/PULL_REQUEST_TEMPLATE/new_model.md?plain=1

It would help me verify the implementation much quicker.