Add offline model support: --ssl-repo-path and --download-models

[chanshing]

Users in offline or restricted environments have no way to pre-download all model artifacts. The .joblib.lzma files are fetched lazily on first run, and the ssl-wearables torch hub repo is cloned when the SSL model loads. This PR adds --ssl-repo-path to use a local repo copy and --download-models to pre-cache everything ahead of time.

Files Changed

src/stepcount/sslmodel.py

• Add repo_path parameter to get_sslnet(). When provided, loads via torch.hub.load() with source='local', skipping GitHub download and cache lookup
• Document that tag is ignored when repo_path is provided

src/stepcount/models.py

• Add ssl_repo_path parameter to WalkDetectorSSL.__init__()
• Pass it through to get_sslnet() in fit() and load_model()
• Use getattr for backward compat with models serialized before this change

src/stepcount/stepcount.py

• Add --ssl-repo-path CLI argument for specifying a local ssl-wearables repo (SSL only)
• Wire it to model.wd.ssl_repo_path after model loading
• Make filepath positional arg optional (nargs='?') so --download-models works without an input file
• Add --download-models CLI flag and early-exit logic in main()
• Add download_models() function: downloads both ssl and rf .joblib.lzma files with atomic temp-file writes, MD5 verification, 60s timeout; caches ssl-wearables repo; validates --ssl-repo-path if provided

tests/test_sslmodel.py

• Test repo_path uses source='local' and skips cache logic
• Test repo_path=None falls back to existing GitHub/cache logic

tests/test_models.py

• Test WalkDetectorSSL stores and defaults ssl_repo_path

tests/test_stepcount.py

• Test --ssl-repo-path appears in CLI help output
• Test --download-models appears in CLI help output
• Test --download-models works without a filepath arg (mocked)
• Test omitting filepath without --download-models gives an error