fix: stop Enum columns raising TypeError on _enums

[TangoEnSkai]

Context

Any operation touching an Enum column fails:

TypeError: String.__init__() got an unexpected keyword argument '_enums'

Enum subclasses String, so with no colspecs entry of its own it resolves
through the String entry and SQLAlchemy adapts it to DatabricksStringType.
That class is a TypeDecorator, and TypeDecorator.__init__ passes its
arguments straight to its impl (String) — while Enum.adapt() forwards
Enum's internal keyword arguments. String rejects them.

metadata.create_all() on a model with an Enum column cannot run, as reported.

Worth noting where the crash actually lives: Enum("A", "B").compile(dialect)
succeeds and renders STRING. Only dialect_impl() — the path every real
query and DDL run takes — raises. That is why the existing camel_case_type_map
entry for Enum passes today without catching this.

What

• Add DatabricksEnumType, a subclass of sqlalchemy.types.Enum that overrides
  only literal_processor() to escape literals the way DatabricksStringType
  does.
• Register it in colspecs so Enum stops resolving through String.
• Add TestDatabricksEnum covering the adaptation that used to crash, DDL
  compilation, native Python enum.Enum columns, literal escaping parity with
  String, length inference, validate_strings enforcement, and that ordinary
  String columns are unaffected.

Why

Two other approaches were tried first and rejected on evidence:

Filtering the offending keyword out is whack-a-mole. Dropping _enums just
moves the failure to _disable_warnings, and the set of forwarded keywords
belongs to SQLAlchemy, not to us.

Letting Enum fall back to SQLAlchemy's own implementation (mapping
Enum → sqlalchemy.types.Enum) fixes the crash but changes literal rendering:

approach	O'Bri\en renders as
String today (the target behaviour)	'O\'Bri\\en'
fall back to generic Enum	'O''Bri\en' ❌
this PR	'O\'Bri\\en' ✅

The middle row is precisely the single-quote doubling that DatabricksStringType
exists to prevent — fixing one bug by reintroducing another.

Subclassing Enum keeps every behaviour the generic type provides and changes
only the literal rendering. Checked against the SQLite dialect as a reference:

	validate_strings=True, bad value	default
SQLite (reference)	LookupError	passes through
this PR	LookupError	passes through

A TypeDecorator-based variant was also prototyped and discarded: it silently
lost validate_strings enforcement, letting "ZZZ" through where SQLite raises.

Completion Criteria

• Enum columns adapt without TypeError, including the reported
  create_all() path
• Enum literals escape identically to String literals
• Enum validation, length inference and native enum.Enum support preserved,
  matching the SQLite dialect
• Plain String columns unaffected
• Removing only the colspecs entry fails exactly the four new tests that
  depend on it, and nothing else
• pytest tests/test_local (offline modules) — 302 passed
• black --check clean on all three changed files
• CHANGELOG entry under a new # Unreleased section
• Commit signed off (DCO)

Note

This touches base.py and CHANGELOG.md, which #74 also touches. The base.py
edits are far apart and merge cleanly; the CHANGELOG entries both open an
# Unreleased section, so whichever lands second needs a one-line rebase. Happy
to do that whenever you'd like.

close #61