Fix successful MLX tokenizer loads

[krystophny]

Summary

• fix load_model_with_fallback() so successful mlx_lm.load() calls return the (model, tokenizer) tuple instead of falling through as None
• add regression coverage for the successful load path

Why this is independently deployable

• pure loader bugfix
• no API, CLI, batching, or protocol behavior change
• useful regardless of whether the Responses work is merged

Related context

This bug sits beneath several other Apple Silicon model-serving paths.

Relevant surrounding work in waybarrios/vllm-mlx:

• #127 merged broader Qwen3.5 text support and streaming fixes: Add Qwen3.5 model support (text-only) and fix reasoning+tool streaming waybarrios/vllm-mlx#127
• open hybrid/cache/runtime work continues in #144, #160, #165, #183, and #194: fix: handle 3D KV tensors in prefix cache for Qwen3.5 models waybarrios/vllm-mlx#144 fix: pass size to ArraysCache in BatchMambaCache for Qwen3.5 hybrid models waybarrios/vllm-mlx#160 fix: MLLM continuous batching for hybrid models waybarrios/vllm-mlx#165 fix: compatibility with mlx-lm 0.31.x (prompt_checkpoints tuple) waybarrios/vllm-mlx#183 fix: unpack prompt_checkpoints in _chunked_next for mlx-lm >= 0.31.0 waybarrios/vllm-mlx#194
• our fork’s separate hybrid-cache follow-up lives in krystophny/vllm-mlx#6: Fix Qwen3.5 hybrid paged cache reconstruction #6

This PR deliberately stays below all of that and only fixes the successful return path.

Validation

• PYTHONPATH=/Users/ert/code/vllm-mlx /Users/ert/code/.venv/bin/python -m pytest tests/test_tokenizer_utils.py -q
• python3 -m compileall vllm_mlx

What could still improve

• broader loader-path coverage for strict/strict-false fallbacks and hybrid model families
• explicit end-to-end smoke tests for each benchmark model alias used by FortBench