feat(hub): add local-directory mode for offline model loading

[Niexiawei]

Summary

• Add hub.NewLocal(dir) / Repo.WithLocalDir(dir), letting a *hub.Repo read model files
  directly from a plain local directory (e.g. output of git clone or
  huggingface-cli download --local-dir) instead of downloading from HuggingFace Hub.
• No network access is made in local-directory mode: DownloadInfo scans the directory to
  build the file listing, and DownloadFile(s) resolve repo-relative names to paths inside it.
  FileURL returns an explicit error since there is no remote URL in this mode.
• Every existing consumer of *hub.Repo — models/transformer, models/safetensors,
  models/gguf, models/sam2, tokenizers/* — works unchanged, since they only depend on
  DownloadInfo/DownloadFile(s)/HasFile/IterFileNames/IterFileInfos, all of which are
  branched internally in hub.
• Design rationale, alternatives considered, and known limitations (e.g. datasets is out of
  scope, standard-cache offline mode is a separate follow-up) are written up in
  docs/local-model-loading-plan.md.

Test plan

• go build ./... and go vet ./... pass
• go test ./hub/... passes, including 12 new tests in hub/local_test.go covering:
  forward-slash file names, .git/.cache exclusion, file sizes, HasFile,
  DownloadFile success/missing-file/path-traversal-safety, FileURL erroring in
  local mode, missing-directory error, and that scanning never mutates the local directory
• go test ./... -short passes across the repo (one unrelated pre-existing failure in
  models/transformer on Windows, caused by an XLA PJRT auto-installer symlink-permission
  issue, not by this change)
• Manually verified end-to-end against a real model (BAAI/bge-small-zh-v1.5) in a
  downstream consumer project: pointed it at a local snapshot directory via hub.NewLocal,
  ran embedding + similarity search with zero network calls, and got correct results

[janpfeifer]

Thank you @Niexiawei !

I removed the AI planning file, I hope you don't mind. Otherwise it looks great ... and I was meaning to do this for a while now, I'm so happy you already did it.

Merging here.

[janpfeifer]

Btw, @Niexiawei , inspired by your PR, I went further in PR #54 and added the following:

• Repo.Save(): save a repo to local disk.
• Repo.DeleteCache(): erase a repo from cache (save space).
• NewEmbed() load a repo that is embedded directly into the Go program (using //go:embed <dir>). This way one can deploy the model along with the go binary, in the same file.
• The provided cmd/hubinfo CLI tool now also accepts:
  • -save: to save a repo to a local directory, so it can be used with NewLocal() (an alternative to huggingface-cli)
  • -delete_cache: free up the space used in cache for a repo.
  • -local: show info about a repo in a local directory.

I also updated the README.md with examples of using local directories.

Check it out!