Bump mlx-swift-lm to 3.x (3.31.3)

[noorbhatia]

Resolves #157

mlx-swift-lm 3.x decouples MLXLMCommon from the Hugging Face Hub and
swift-transformers' tokenizer: it now defines its own Tokenizer,
Downloader, and TokenizerLoader protocols, and consumers must supply
concrete implementations. Adopt the in-repo MLXHuggingFace macros to
restore Hub-based loading with minimal churn, keeping the existing
ModelContext-based backend intact.

Package.swift:

• Bump mlx-swift-lm to from: "3.0.0" (resolves 3.31.3; pulls mlx-swift
  0.31.x).
• Add swift-huggingface (0.9.0) dependency.
• Add MLX-trait products: MLXHuggingFace (macros), HuggingFace
  (HubClient), and swift-transformers' Tokenizers (AutoTokenizer).

MLXLanguageModel.swift:

• Drop import Hub; add MLXLLM, MLXHuggingFace, HuggingFace imports.
  MLXLLM is now imported so MLXLLM.TrampolineModelFactory is linked
  alongside MLXVLM's for ModelFactoryRegistry resolution.
• Replace the removed loadModel(hub:id:) / loadModel(directory:) with
  the 3.x loadModel(from:using:id:) / loadModel(from:using:), expanding
  the #hubDownloader() and #huggingFaceTokenizerLoader() macros.
• Rename Tokenizer.decode(tokens:) -> decode(tokenIds:) (3 sites) and
  qualify any Tokenizer -> any MLXLMCommon.Tokenizer to disambiguate
  from Tokenizers.Tokenizer.

BREAKING (MLX trait only): MLXLanguageModel.hub and its hub: init
parameter change from HubApi? to HubClient?. Code passing or reading a
custom hub must switch import Hub + HubApi(...) to import HuggingFace

• HubClient(...). MLXLanguageModel(modelId:) and other default
  constructions are unaffected; non-MLX consumers see no change.