feat: offline cache probe paired with download helper

[wavekat-eason]

Summary

Adds an offline, pure-filesystem cache probe so consumers can answer "is this preset already on disk?" without touching the network.

• download::is_repo_cached(repo_id, files) -> bool — generic, mirrors the existing download_files_with_progress(repo_id, files, dest_dir, on_progress) signature.
• ModelPreset::is_cached(&self) -> bool — convenience method for the sherpa-onnx backend's preset constants.

Why

hf-hub 0.5 has no HF_HUB_OFFLINE support and no public lookup-only download API. The natural "try to construct the backend and see if it succeeds" probe therefore ends up firing a full HuggingFace metadata + download cycle on every cold-cache call — a UI that uses it as a quick "is the model ready?" check ends up silently downloading ~100 MB in the background and never showing its "Download model" CTA. Consumers that want to stay offline during a UI render need a pure-filesystem alternative.

How

The probe walks the on-disk layout hf-hub writes:

<cache_root>/models--<owner>--<repo>/snapshots/<commit>/<file>

…and returns true when at least one snapshot dir contains every requested file. The cache root is resolved exactly the way hf-hub 0.5's Cache::from_env does: honor HF_HOME (appending hub), otherwise default to \$HOME/.cache/huggingface/hub. USERPROFILE is also accepted as a Windows fallback for HOME.

The implementation is split into:

• pub fn is_repo_cached — public entry point, looks up cache_root from the environment.
• fn hf_hub_cache_root — env resolution, factored out for clarity.
• fn snapshot_satisfies_files(cache_root, repo_id, files) — pure logic, factored out so the unit test can drive it against a temp dir without touching the real cache or polluting process-wide env vars.

ModelPreset::is_cached is a one-liner over is_repo_cached(self.model_id, &self.files()), paralleling the existing download_preset_with_progress wrapper.

Tests

Two new unit tests in download::tests:

• snapshot_probe_matches_hf_hub_layout — builds the cache layout step by step in a temp dir and asserts (a) false for missing repo / empty snapshot dirs / partial files and (b) true once every requested file is present in at least one snapshot. Also covers a second, half-complete snapshot not invalidating a first complete one.
• snapshot_probe_flattens_slashes_in_repo_id — guards against a silent regression where a slash-bearing repo id (the common case) wouldn't be flattened to <owner>--<repo> and the probe would always look in a non-existent directory.

Pure-fs, no network, no real models needed — runs under the existing make ci matrix.

Test plan

• make test (9 passed, 0 failed)
• make lint (clippy clean)
• make ci (fmt + clippy + test + doc-test all green)

🤖 Generated with Claude Code