[Fix] Use SQLite for default MLflow tracking

[appleweiping]

Motivation

MLflow 3.14 rejects filesystem tracking stores unless the compatibility escape hatch is enabled. This makes the default MLflowVisBackend initialization fail. Fixes #1687.

Modification

When tracking_uri is not supplied, MLflowVisBackend now stores metadata in a SQLite database at save_dir/mlflow.db and artifacts under save_dir/artifacts. Explicit tracking_uri and artifact_location values keep their existing behavior.

The regression test verifies that the default tracking URI is SQLite and that the database is created. It uses a non-persistent SQLAlchemy pool during tests so Windows can remove the temporary database.

BC-breaking

The default local metadata store changes from the legacy MLflow filesystem format to SQLite. Existing filesystem runs are not migrated automatically. They can be migrated with MLflow's migrate-filestore command. Explicit tracking URIs are unchanged.

Validation

• Reproduced the failure with MLflow 3.14.0
• MLflowVisBackend tests: 8 passed
• Visualization backend test file: 37 passed and 14 skipped before unrelated DVCLive tests failed because optional dvclive and pygit2 packages were not installed
• flake8, isort, YAPF, docformatter, and pyupgrade passed for the changed files

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[appleweiping]

The CircleCI lint job fails during pre-commit initialization, before it checks the changed files. Its Python 3.7 environment resolves a pre-commit version that rejects the hook manifest stage named pre-commit. The changed files pass flake8, isort, and YAPF locally.