refactor: consolidate split logic in base class

[ziv-lazarov-nagish]

Summary

• Add split_bucket() and assign_split() to common.py — deterministic hash-based split assignment
• Add _init_split_tracking(), _track_and_filter(), get_split_manifest() to BaseSegmentationDataset — removes duplication across all dataset classes
• DGS: keep fixed splits.json (research-standard split), migrate to base class tracking infra and pathlib
• Platform: use assign_split() + base class helpers, remove local _split_bucket
• Remove unused collate_fn/DataLoader imports and duplicate manifest collection in train.py

Depends on #19.

Changed files

• datasets/common.py — split_bucket, assign_split, base class split tracking
• datasets/dgs/dataset.py — use base class tracking while keeping fixed splits.json, migrate to pathlib
• datasets/annotation_platform/dataset.py — use assign_split + base class helpers
• train.py — remove unused imports and duplicate manifest
• tests/ — update imports and dummy dataset

Test plan

• ruff check . passes
• pytest passes (61 tests)
• 1-epoch DGS training smoke test — splits match original (545 train, 9 dev, 14 test)
• 1-epoch dgs,platform training smoke test — both datasets load correctly

[AmitMY]

the split we use for dgs is a fixed one, used in research (not only our research)
using the split.json is preferred here - because it allows us to compare to our previous work, and to other's work as it comes. as i understand here we would create a new split

[ziv-lazarov-nagish]

the split we use for dgs is a fixed one, used in research (not only our research) using the split.json is preferred here - because it allows us to compare to our previous work, and to other's work as it comes. as i understand here we would create a new split

correct. i was thinking on a single way to split it across all SegmentationDataset classes for reproducibility. i'll make an exception for the DGS one