Add dense qwen3_5 support for learned quantization

[iamwavecut]

Summary

This PR adds learned quantization support for dense qwen3_5 text models and hardens DWQ failure handling.

Changes:

• Enable AWQ traversal for dense qwen3_5 hybrid layers:
  • full attention blocks with self_attn.{q,k,v,o}_proj
  • linear-attention blocks with linear_attn.in_proj_*
  • dense MLP projections
• Avoid the qwen3_5 inference-only gated-delta CustomKernel during dynamic quantization sensitivity estimation.
• Add dynamic_quant calibration controls for small local runs:
  • --num-samples
  • --sequence-length
• Add DWQ KL loss implementation selection:
  • --kl-loss-impl auto|metal|mlx
• Add DWQ target metadata validation before training.
• Add DWQ shape and finite-value checks so NaN/Inf losses or weights fail before saving a corrupted checkpoint.
• Add unit tests and an opt-in slow qwen3_5 quantization smoke test.

This intentionally focuses on dense qwen3_5 models. MoE support is left out of scope.

Motivation

Plain conversion and GPTQ already work for qwen3_5 models, but other learned quantization paths had gaps:

• AWQ rejected qwen3_5 as unsupported.
• Dynamic quantization could hit [Primitive::vjp] Not implemented for CustomKernel on qwen3_5 linear-attention paths.
• DWQ could fail on Apple Silicon devices where the Metal KL-loss kernel configuration is not usable, and unstable runs could continue far enough to save invalid checkpoints.

Testing

uv run python -m unittest tests.test_losses tests.test_quantization -v

uv run python -m unittest discover tests -v

uv run pre-commit run --files \
  mlx_lm/LEARNED_QUANTS.md \
  mlx_lm/quant/awq.py \
  mlx_lm/quant/dwq.py \
  mlx_lm/quant/dynamic_quant.py \
  mlx_lm/tuner/losses.py \
  tests/test_losses.py \
  tests/test_quantization.py \
  tests/test_qwen3_5_quantization_slow.py

Opt-in slow smoke test on Qwen/Qwen3.5-0.8B:

RUN_SLOW_QWEN35_QUANT_TESTS=1 \
QWEN35_QUANT_TEST_MODEL=/path/to/Qwen3.5-0.8B \
uv run python -m unittest tests.test_qwen3_5_quantization_slow -v

Local smoke checks also verified that AWQ, dynamic quantization, and DWQ start and save loadable artifacts for the downstream finetuned model LakoMoor/QClaw-4B. GPTQ was checked to reach Hessian collection and quantization startup as an existing baseline.