refactor: shared get_dataloader between train and evaluate

[ziv-lazarov-nagish]

Summary

• Move get_dataloader from train.py to datasets/common.py with explicit args parameter
• evaluate.py uses shared get_dataloader instead of manual build_datasets + DataLoader
• CI: run lint and test workflows on PRs targeting nagish branch

Changed files

• datasets/common.py — add get_dataloader() with **augment_overrides support
• train.py — remove local get_dataloader, import from common
• evaluate.py — use get_dataloader instead of manual construction
• datasets/__init__.py — export get_dataloader
• .github/workflows/test.yaml / lint.yaml — add nagish to PR branch triggers

Test plan

• ruff check . passes
• pytest passes (61 tests)

[AmitMY]

actually, shouldn't some of it go into main?

[ziv-lazarov-nagish]

actually, shouldn't some of it go into main?

it will, i was planning to cherry-pick the commits from nagish later instead of merging into main and then update nagish.

[AmitMY]

my approach would have been to update main, and then nagish
but you can also update nagish then ask claude to make a PR to main selecting only the non-spepcif nagish changes

[ziv-lazarov-nagish]

my approach would have been to update main, and then nagish but you can also update nagish then ask claude to make a PR to main selecting only the non-spepcif nagish changes

that was my intention since i thought most of our efforts would be in the nagish branch going forward, so treating it as the "main" branch