feat: delegate constraint naming to Databricks

[Tomoscorbin]

Summary

• stop synthesizing primary- and foreign-key constraint names when users omit them
• let Databricks allocate physical names in both CREATE TABLE and ALTER TABLE ... ADD SQL
• treat an explicit name as a creation preference, not ongoing managed state
• match existing constraints by structural definition with ordinary equality
• model catalog-observed constraints separately so their physical names are always available for drops
• update tests and documentation around the final behavior

Behavior

Constraint names now have one narrow meaning:

• no name: Databricks chooses the physical name when creating the constraint
• explicit name: request that physical name when creating the constraint

Once a constraint exists, its name belongs to the catalog. A desired constraint is satisfied by an observed constraint with the same definition, regardless of either name. Changing only primary_key_name or ForeignKey(name=...) is therefore a no-op.

Structural drift still produces a drop and an add. The drop uses the concrete observed name; the add preserves the user's optional creation preference. Existing tables using the old engine-generated naming convention converge without DDL churn.

An explicit name collision during creation remains a Databricks execution error because constraint names share a schema-wide namespace.

Design

• PrimaryKeyConstraint and ForeignKeyConstraint represent desired declarations and carry an optional name
• ObservedPrimaryKeyConstraint and ObservedForeignKeyConstraint require the catalog name, making unnamed observed state unrepresentable
• constraint equality and hashing use only the relational definition
• primary-key reconciliation is a direct desired_key == observed_key comparison
• foreign-key reconciliation pairs structural matches one-to-one, then drops unmatched observations before adding unmatched declarations
• optional-name branching is confined to declaration validation, desired-model construction, reporting, and SQL rendering
• drop actions carry only the physical name required by Databricks

Validation

• uv run pytest -q: 1,244 passed, 79 deselected; 97.21% coverage
• uv run mypy src: passed
• uv run ruff check .: passed
• uv run ruff format --check .: passed
• uv run lint-imports: 7 contracts kept
• uv run --group docs sphinx-build -W -b html docs docs/_build/html: passed
• standalone MyST plan render: passed
• live suite collection: 79 tests collected successfully; credentialed tests not executed locally
• git diff --check: passed