fix: merge user-supplied fit(callbacks=...) with built-in trainer callbacks

[dripston]

Problem

Calling fit(callbacks=[...]), exactly as docs/core_concepts/config_system.md and docs/core_concepts/training_and_evaluation.md document as a supported Lightning passthrough, crashes:

TypeError: lightning.pytorch.trainer.trainer.Trainer() got multiple values for keyword argument 'callbacks'

One of several independent defects filed together in #452 — this PR addresses only the callbacks= collision, since the others touch unrelated code paths (seeding, fit-failure state corruption, NODE initialization, predict(device=), profile(dry_run=True)).

Root cause

deeptab/models/_mixins/fit.py's pl.Trainer(...) construction hard-codes callbacks=[early_stop_callback, checkpoint_callback, ModelSummary(...)] and then spreads **trainer_kwargs after it. Any callbacks= in trainer_kwargs collides with the hard-coded keyword argument. logger= a few lines below already handles this correctly by popping the user's value out of trainer_kwargs and merging it in; callbacks had no equivalent handling, so there is no way to attach a custom callback (LR monitors, Optuna pruning, gradient-accumulation schedulers) despite the docs saying otherwise.

Fix

callbacks=[
    early_stop_callback,
    checkpoint_callback,
    ModelSummary(max_depth=2),
    *trainer_kwargs.pop("callbacks", []),
],

Mirrors the existing logger=trainer_kwargs.pop("logger", ...) pattern immediately below it. When no callbacks= is passed, behavior is unchanged (empty list unpacks to nothing). When one is passed, it's appended after the built-ins instead of colliding with them.

Testing

• Added test_fit_accepts_user_supplied_callbacks in tests/test_models.py, following the existing MLPRegressor/regression_data/FIT_KWARGS fixtures used throughout that file. It fits with an explicit callbacks=[LearningRateMonitor()] and asserts both the user's callback and the built-in callbacks are present on model._trainer.callbacks.
• python3 -m py_compile passes on both changed files.
• Manually simulated the list-merge logic in isolation (pure Python, no Lightning dependency) against both the "user passes callbacks=" and "user passes nothing" cases — confirmed no collision and no change to default behavior.
• Could not run the actual test suite in this environment (lightning/torch are not installed here) — please run pytest tests/test_models.py -k callbacks in CI/a configured dev environment to confirm.

Addresses #452 (callbacks portion only — not closing, since the issue bundles 6 additional unrelated defects)

🤖 Generated with Claude Code