Add clean cache option to forecaster

[elmartinj]

Summary

Adds an optional clean_cache parameter to TimeCopilotForecaster.

When enabled, the forecaster runs cache cleanup after each model call inside the shared _call_models() loop. This helps memory-heavy workflows, especially when running multiple foundation models sequentially.

Changes

• Added clean_cache: bool = False to TimeCopilotForecaster.
• Added _clean_model_cache() helper.
• Cleanup currently runs:
  • gc.collect()
  • torch.cuda.empty_cache() when Torch is installed and CUDA is available
• Integrated cleanup after each model forecast/cross-validation call.
• Added a unit test confirming cleanup runs once per model when clean_cache=True.

Validation

uv run pytest -n 2 tests/test_forecaster.py -q
# 12 passed

uv run pre-commit run --all-files
# passed

[AzulGarza]

thanks @elmartinj! lgtm.:)