Improve DLC project import integration

[talmo]

Migrating from talmolab/sleap#734 since DLC I/O now lives in sleap-io
(sleap_io/io/dlc.py).

Current DLC import (load_dlc) reads CSV-based annotations into Labels, but
several pieces of DLC project metadata are still missing for a complete import:

1. Training set splits: DLC stores train/val/test splits in
   Documentation_data-*.pickle (and related metadata files). sleap-io
   doesn't currently surface these — they'd map most naturally to a future
   per-frame split annotation or to a sidecar config consumed by sleap-nn.

2. Cropping / ROI: DLC supports ROI cropping during labeling/training.
   sleap-io has no ROI primitive yet, so this is blocked on adding ROI
   support to the data model.

3. Link single images back to original videos: DLC exports labeled
   frames as folders of images under labeled-data/<video_name>/. Today
   sleap-io creates a Video.from_filename(image_files) per folder; it
   could additionally parse the DLC config.yaml (which lists
   video_sets) to point back to the original .mp4/.avi files when
   available.

4. Skeleton edges: _parse_multi_animal_structure and
   _parse_single_animal_structure (sleap_io/io/dlc.py:196-257) only
   build Skeleton(nodes=...) and never populate edges. DLC's
   config.yaml defines skeleton: as a list of edges — these should be
   parsed and added to the imported Skeleton.

Originally filed as talmolab/sleap#734 (2022). Related: talmolab/sleap#606.

This will likely be broken down into separate PRs / sub-tasks per item.

[talmo]

Items 1, 3, and 4 are implemented and merged in #450:

• Skeleton edges (4) — parsed from config.yaml skeleton: onto the imported Skeleton (edges referencing unlabeled bodyparts are dropped with a warning; name-based, never injects nodes).
• Source-video links (3) — each labeled-data/<video>/ image folder is linked to its original video from video_sets via Video.source_video, matched by filename stem.
• Train/test splits (1) — load_dlc_splits(config) -> LabelsSet reconstructs DLC's globally sort_index()-merged frame order (with its skip-rules) and maps the Documentation_data-*.pickle indices to "train"/"test".

Also added load_dlc_project() (whole-project load + merge), a config= kwarg on load_dlc (auto-discover / explicit / False), and load_file routing for DLC project directories / config.yaml.

Item 2 (cropping) — future scope

Refining the original framing: sleap-io already has an ROI data-model primitive. What it lacks is the ability to apply an ROI as a virtual, read-time crop on a Video — which is exactly how DLC does cropping. DLC's VideoReader holds a bbox and slices frame[y1:y2, x1:x2] on each read (deeplabcut/utils/auxfun_videos.py), and video_sets[...].crop is just that per-video ROI (default = full frame). Frames extracted for labeling are read through this virtually-cropped reader, so labeled-data/<video>/ images are the cropped region and labels live in cropped-frame coordinates.

Implications for a future implementation here:

• The missing capability is virtual ROI-crop-on-read for Video, not a new ROI primitive. Once that exists, importing video_sets[...].crop (as an ROI applied to the source video) becomes natural.
• As shipped in feat(io): import DLC project metadata — edges, source videos, splits (#424) #450, source_video points at the original uncropped video, so for cropped projects there is an (x1, y1) coordinate offset between the labels (cropped-frame coords) and the source video. Virtual cropping is what reconciles the two. This caveat is documented in the DeepLabCut format docs.

Leaving this issue open to track item 2; not filing a separate tracking issue for now.

[talmo]

Item 2 (virtual ROI cropping) is now scoped and tracked in #451 (CropVideoBackend — virtual on-read crop with a /video_crops SLP dataset, backward-compatible). Items 1, 3, and 4 shipped in #450. Closing this in favor of #451.