refactor: drop hive, databricks, and redshift connectors

[mplatzer]

Pull Request

Stacks on top of #720 (feat/local-first-major-release). Drops the three most complex data connectors so the LOCAL-first SDK only ships connectors we can actually maintain and test.

Changes

• Delete the connector modules:
  • mostlyai/sdk/_data/db/hive.py
  • mostlyai/sdk/_data/db/databricks.py
  • mostlyai/sdk/_data/db/redshift.py
• Remove ConnectorType.{hive, databricks, redshift} from mostlyai/sdk/domain.py (enum + docstring entries) and the matching mostly.connect examples in mostlyai/sdk/client/api.py.
• Drop the databricks, hive, and redshift extras from pyproject.toml and refresh the README's connector list / install example.
• Remove DatabricksContainerParameters plus the Hive-only keystore / keystore_password fields and all Kerberos fields from mostlyai/sdk/_data/metadata_objects.py. The shared ca_certificate field is kept (still used by AWS S3 and Postgres SSL).
• In mostlyai/sdk/_data/db/base.py, remove the kerberized() context manager, all Kerberos init args / attributes / helpers, the KRB5_CONF_TEMPLATE, and now-unused base64 / hashlib / os / subprocess imports. use_sa_engine now only wraps use_ssh_tunnel() and use_ssl_connection().
• Delete mostlyai/sdk/_data/util/kerberos.py and its unit tests (tests/_data/unit/util/test_kerberos.py); these were used exclusively by the Hive connector.
• Refresh uv.lock accordingly.

Why this change?

Per the LOCAL-first refactor goal, we want a connector surface that we can credibly support. These three are by far the heaviest:

• Redshift ships its own DefaultDialect-derived SQLAlchemy dialect and a hand-rolled DBAPI wrapper, plus bespoke information_schema queries, case-insensitivity workarounds, and parameter-limit chunking. It's the largest single connector in the repo.
• Databricks owns a non-standard write path (parquet → temp Volume → COPY INTO), Azure ClientSecretCredential service-principal auth, and a fragile error-message-keyword mapper in is_accessible().
• Hive monkey-patches ImpalaDialect and HiveDialect for SA 2.0 compatibility and was the sole driver of the Kerberos plumbing in SqlAlchemyContainer and the kerberos/pyhive/impyla extras.

None of these had end-to-end coverage in this repo (only SQLite is covered by tests/_local/end_to_end/test_connector.py), so dropping them removes both the most code and the most untested code.

The remaining DB connectors (postgresql, mysql, mariadb, mssql, oracle, snowflake, bigquery, sqlite) are thin SQLAlchemy wrappers around mostlyai/sdk/_data/db/base.py, and all file connectors are unaffected.

Testing

• uv run python -m compileall -q mostlyai/sdk — clean.
• uv run pytest tests/test_domain.py tests/_local/unit/ tests/client/unit/ tests/_data/unit/db/ tests/_data/unit/util/ tests/_data/unit/file/ — 177 passed, 1 skipped.
• uv run pytest tests/_local/end_to_end/test_connector.py — 4 passed (SQLite read / write / delete / query).
• uv run ruff check on the modified files — all clean. (Pre-existing lint warnings in unrelated files under mostlyai/sdk/_local/execution/ and mostlyai/sdk/_data/{auto_detect,non_context}.py are inherited from the base branch and not touched here.)

Additional Notes

• Public API impact: ConnectorType loses the HIVE, DATABRICKS, and REDSHIFT members. Existing connectors of those types stored on disk would no longer load — this is expected for a local-first major-release branch.
• The shared ca_certificate field is intentionally retained on SqlAlchemyContainerParameters and SslCertificates; only the Hive-specific keystore/keystore_password fields are removed.