fix(qwen3): support rope_parameters config

[daniellaah]

• ☑️ I understand it is strictly prohibited to use AI to write PR description
• AI usage disclosure: Use Claude Code for testing and review

Summary

Qwen3 checkpoints produced by transformers 5.x releases (e.g. dinghar/qwen3-0.6b-pangram-sft) fail to load in mlx_lm.models.qwen3 because ModelArgs cannot load rope_theta correctly.

The checkpoint's config.json provides:

"rope_parameters": {"rope_theta": 1000000, "rope_type": "default"}

with no top-level rope_theta field. But ModelArgs.rope_theta is a required field with no default, so loading fails with:

TypeError: ModelArgs.__init__() missing 1 required positional argument: 'rope_theta'

In addition, rope_parameters may contain RoPE configuration such as rope_type, factor, and original_max_position_embeddings. ModelArgs currently ignores rope_parameters, so this configuration can be lost without raising an error because rope_scaling is optional.

Fix

ModelArgs:

• switch rope_theta from a required field to an optional one so ModelArgs can get it from rope_parameters
• add an optional rope_parameters: Optional[dict].
• resolve rope_theta from rope_parameters when the top-level value is absent, and use rope_parameters as the fallback RoPE scaling configuration when rope_scaling is absent.

Changes are backward compatible: the existing test_qwen3 (flat rope_theta, no rope_parameters) still passes unchanged.

Testing

• Added test_qwen3_rope_parameters:
  • nested config works
  • top-level fields take precedence
  • raises ValueError when rope_theta is absent from both
• Confirmed the new test fails with the exact TypeError above against unpatched main, and passes after the fix.
• Verified against the real checkpoint end to end: dinghar/qwen3-0.6b-pangram-sft now loads correctly.
• Full tests/test_models.py suite passes locally.
• Passed uvx pre-commit run --all

Fixes #1563

[michalk8]

LGTM, thank you!