fix(postgres): correct SQL for modifying a column to :identity

[nabsei]

Fixes #4751.

What

column_change/2 for {:modify, name, type, opts} reused column_type/2 as-is for the ALTER COLUMN ... TYPE ... clause. For type == :identity, that function produces "bigint GENERATED BY DEFAULT AS IDENTITY", which is valid inside a CREATE TABLE/ADD COLUMN column definition, but PostgreSQL rejects it combined with ALTER COLUMN ... TYPE:

ALTER TABLE "my_table" ALTER COLUMN "field" TYPE bigint GENERATED BY DEFAULT AS IDENTITY
** (Postgrex.Error) ERROR 42601 (syntax_error) syntax error at or near "GENERATED"

The identity clause must instead be added via a separate ALTER COLUMN ... ADD GENERATED BY DEFAULT AS IDENTITY action within the same ALTER TABLE statement.

Fix

• modify_column_type/2 emits just the base type (column_type_name/2) for the TYPE clause when modifying to :identity, leaving column_type/2's behavior for CREATE TABLE/ADD COLUMN untouched.
• New modify_identity/3 emits the separate ALTER COLUMN ... ADD GENERATED BY DEFAULT AS IDENTITY(...) clause, appended after the existing null/default/collation clauses.
• The sequence-option logic (start_value/increment) is extracted into identity_generated_expr/1 and shared between the CREATE TABLE path and the new ALTER path, unchanged in behavior.

Verification

• Full test suite: 683 passed, no regressions (confirmed pre-existing unrelated warnings/formatting diffs are identical on unmodified master).
• Added two tests in test/ecto/adapters/postgres_test.exs covering {:modify, :field, :identity, []} and with start_value/increment options, matching the exact SQL PostgreSQL requires.
• Ran the actual generated SQL against a real local PostgreSQL 16 instance (not just string-matched): confirmed both statements execute successfully, the column becomes a working identity column, and INSERT ... RETURNING shows the configured start_value is honored. Also confirmed the PostgreSQL requirement that the column be NOT NULL before adding an identity, and that this composes correctly with Ecto's existing null: false modify option on a nullable column.

---

This PR was prepared with the assistance of Claude Code (Anthropic's CLI-based coding agent). The root-cause analysis, fix design, and verification steps above were reviewed and confirmed by me before submitting.