feat: Huggingface

[jtschwar]

Summary

Adds support for pretrained checkpoints hosted on the Hugging Face Hub, restructures
membrane extraction into a general extract command group, and finishes migrating the
CLI to URI-based flags (--tomo-uri, --seg-uri, --target-uri) across the remaining
commands.

Changes

Hugging Face Hub checkpoints

• New octopi/utils/hub.py: resolves a checkpoint alias (e.g. tomogram-boundary) against
  the biohub/octopi Hub repo, downloading weights + config on demand and caching locally.
  Falls back to treating --model-weights as a local path when it isn't a known alias.
• octopi segment / octopi.workflows.segment: --model-config is now optional — omit it
  when --model-weights is a Hub alias, since its config is bundled and downloaded
  automatically. Still supports comma-separated paths for model-soup ensembles.
• Adds huggingface_hub as a dependency (pyproject.toml).

extract command group

• Replaces the standalone membrane-extract command with octopi extract, a group with
  two subcommands:
  • extract mb-picks — the existing membrane-proximity pick splitting (unchanged logic,
    moved under the group).
  • extract seg (new, octopi/extract/segmentations.py) — isolates a single object's
    binary mask from a multi-class segment prediction, using the inference log written by
    segment to recover voxel size and label mapping.

URI-based CLI cleanup

• octopi segment: finishes the URI migration — takes --tomo-uri (algorithm@voxel_size)
  and --seg-uri (name:user_id/session_id) directly; URI parsing now happens once inside
  workflows.segment() instead of in the CLI layer.
• octopi.workflows.segment() signature changed to (config, model_weights, model_config, tomo_uri, seg_uri, ...) — parses both URIs internally.
• pytorch/segmentation.py renamed to pytorch/inference.py (predictor classes only,
  no behavior change) to disambiguate from the new extract/segmentations.py.
• Removed the dead nnUNet CLI group reference from entry_points/groups.py.

MCP server

• Updated tool instructions/help text for the new extract group and optional
  --model-config.
• get_command_help now supports group subcommands (e.g. "extract seg").
• Added guidance to default to CLI commands and only suggest the octopi.workflows Python
  API when the user is clearly scripting/notebooking, not running a one-off command.

Docs

• Updated docs/api/inference.md, docs/user-guide/inference.md, docs/user-guide/labels.md,
  and others for the new checkpoint-alias workflow and extract group.