Try local_files_only first in load() to skip avoidable HF Hub round-trips

[Vlor999]

Closes #1570.

Summary

_download() (used by both load() for model weights and load_tokenizer() for tokenizer files) called huggingface_hub.snapshot_download() without local_files_only=True. Even with a complete local cache, snapshot_download performs an API round-trip (repo_info/model_info) to check for updates before touching the cache. Since load() calls _download() twice per invocation (weights + tokenizer files), this cost was paid twice on every single call — for the CLI (mlx_lm.generate, mlx_lm.chat, mlx_lm.server) and every library user, not just first-time downloads.

This PR makes _download() try snapshot_download(..., local_files_only=True) first (near-instant when the cache is complete) and only fall back to the network-enabled call on LocalEntryNotFoundError (the genuine first-download / incomplete-cache case). ModelScope users are unaffected — left on the existing always-network path since I couldn't validate local_files_only semantics for that backend.

Measurements (M4 Pro, 5 runs/config, mean ± stdev)

Model	before	after	change
Qwen2.5-0.5B-Instruct-4bit	0.355 ± 0.030 s	0.178 ± 0.003 s	-49.8%
Llama-3.2-1B-Instruct-4bit	0.592 ± 0.042 s	0.412 ± 0.002 s	-30.5%
gpt-oss-20b-MXFP4-Q8	3.765 ± 0.062 s	3.544 ± 0.052 s	-5.9% (weight loading dominates at this size)

For an isolated cache hit, snapshot_download(..., local_files_only=True) on these repos takes 0.4-1.6ms, confirming the ~200-230ms per call removed was pure avoidable network latency rather than any real cache-validation work.

Note on load_tokenizer: after this fix, the new dominant cost of load() (profiled with cProfile) is transformers.AutoTokenizer.from_pretrained itself (Tokenizer.__setstate__ + a copy.deepcopy of the tokenizer object, ~280ms combined for Llama-3.2-1B). That's upstream transformers/tokenizers code, out of scope for this PR, but worth flagging for anyone looking at load() latency next.

Test plan

• python -m pytest tests/test_utils.py tests/test_generate.py tests/test_tokenizers.py tests/test_prompt_cache.py — 62 passed, 6 subtests passed, no changes needed.
• black --check mlx_lm/utils.py — clean on the diff (an unrelated pre-existing block elsewhere in the file would reformat under a newer black, untouched here).
• Verified the fallback path still works for a repo requiring an actual download (raises LocalEntryNotFoundError on cache miss, falls through to the network-enabled call).

[Vlor999]

Superseded by #1650, which rebuilds this on current main and adds a measurement (147 ms -> 0.26 ms per call) plus tests. Closing to keep the queue clean.