Add Qwen3.5 0.8B and 2B variants

[Lokimorty]

Adds two new Qwen3.5 model options to the in-app model catalog: 0.8B 4bit and 2B 4bit. Updates include model definitions, model list wiring, reasoning classification, and memory size metadata used by onboarding filters. README supported-model links are updated to match the new options. This also switches package wiring from mlx-swift-examples to mlx-swift-lm (and aligns swift-transformers to 1.1.x) so qwen3_5 model_type is supported at runtime instead of failing with unsupportedModelType.

[missuo]

Your PR has issues. Did you try to build and use the model? It will report an error: Failed: unsupportedModelType("qwen3_5")

[missuo]

I've tested it, and the 9B model works. Can you add the 9B model?

[Lokimorty]

Adding just the model entries was not enough for Qwen3.5 to work at runtime. To support them end-to-end, I switched package from mlx-swift-examples to mlx-swift-lm for MLXLLM/MLXLMCommon. With the old package path, runtime failed with: unsupportedModelType("qwen3_5")

In the initial commit I scoped it to model list changes assuming that was a known issue.

New commits are 2 dependency fixes:

1. Bump swift-transformers to 1.1.x to satisfy the mlx-swift-lm dependency graph.
2. Add direct mlx-swift / MLXRandom linkage because LLMEvaluator imports MLXRandom; without explicit linkage, Xcode fails with missing module dependency errors.
3. Also added 9B model by missuo's request

So the extra changes are to unblock build + runtime support for the new Qwen3.5 variants.

[missuo]

1. Bump swift-transformers to 1.1.x to satisfy the mlx-swift-lm dependency graph.

You should remove the swift-transformers dependency, as mlx-swift-lm already includes it.