Fix TFT fit_from_dataset future covariate validation

[kiwoongyoon]

Fix TFT fit_from_dataset future covariate validation

Summary

• Raise the existing TFT future-covariate ValueError during _create_model() when a custom training dataset has no future covariates and add_relative_index=False.
• Add a regression assertion covering TFTModel.fit_from_dataset() with SequentialTorchTrainingDataset.

Test evidence

• UV_PROJECT_ENVIRONMENT=/tmp/gitcontribute-darts-issue-3008-venv uv run --with '.[torch]' --with pytest python -m pytest darts/tests/models/forecasting/test_TFT.py::TestTFTModel::test_future_covariate_handling -q passed: 1 passed, 4 warnings in 10.87s.
• git diff --check passed.
• python -m py_compile darts/models/forecasting/tft_model.py darts/tests/models/forecasting/test_TFT.py passed.

Risks or notes for maintainers

• The error message is duplicated with the existing _build_train_dataset() validation to keep this patch narrowly scoped.
• Full test suite was not run; the targeted TFT regression test passed with project torch dependencies installed through uv.

Fixes #3008