fix: add the two missing @reflection.cache decorators (get_foreign_keys, get_columns)

[TangoEnSkai]

Context

SQLAlchemy threads an info_cache dict through every Inspector call so that
reflecting the same table repeatedly costs one round-trip rather than one per
call. A dialect method only participates in that cache when it is decorated with
@reflection.cache.

Two methods in DatabricksDialect were missing it — get_foreign_keys() (#72)
and get_columns() (#75). Both issue a warehouse round-trip on every single
call. Their neighbours get_pk_constraint(), has_table(), get_table_names(),
get_view_names(), get_materialized_view_names(), get_temp_view_names(),
get_schema_names() and get_table_comment() are all decorated, which makes
these look like omissions rather than deliberate exclusions.

Inspector.get_columns() shows the kwarg being handed over explicitly:

col_defs = self.dialect.get_columns(
    conn, table_name, schema, info_cache=self.info_cache, **kw
)

It arrived, landed in **kwargs, and was discarded.

Measured against main with the transport stubbed out, three calls sharing one
info_cache:

method	server round-trips	info_cache keys
get_columns()	3	[] — never populated
get_foreign_keys()	3	[] — never populated
get_pk_constraint()	1	populated

For get_columns() the cost is not always a single statement: when
cur.columns() returns an empty list, it follows up with DESCRIBE TABLE EXTENDED to tell a column-less table from a missing one (base.py:156), so an
uncached call on such a table is two round-trips, repeated every time.

Callers that reuse one Inspector across many lookups feel this directly —
Alembic's autogenerate is the common case, as is any long-lived application that
reflects per request.

What

• Add @reflection.cache to get_foreign_keys() and get_columns().
• Add tests/test_local/test_reflection_cache.py covering, for both methods: the
  cache hit, that a direct call passing no info_cache is unaffected (the
  decorator falls straight through when the kwarg is absent, so callers invoking
  the dialect method directly keep today's behaviour), per-table cache keying for
  get_foreign_keys(), and a regression guard on the already-correct
  get_pk_constraint().

get_indexes() is left undecorated on purpose: it returns the EMPTY_INDEX
constant without touching the server, so a cache would buy nothing.

Why

An uncached reflection method is a warehouse query per table per call. The change
is the same mechanism the neighbouring methods already use, and it matches
upstream practice — SQLAlchemy's own SQLite, PostgreSQL and MySQL dialects all
decorate get_columns().

It cannot change results: @reflection.cache is a no-op unless SQLAlchemy itself
supplies info_cache, and the calls it collapses are identical by construction.
Caching is also safe with respect to Inspector._instantiate_types(), which
mutates the returned column dicts in place — it is guarded by
if not isinstance(coltype, TypeEngine), so re-running it over an
already-instantiated cached list is a no-op. This is the same situation the
upstream dialects are in.

Scope note

reflection.cache builds its key from fn.__name__, so this does not make
get_pk_constraint() and get_foreign_keys() share a single DESCRIBE TABLE EXTENDED within one reflection pass — those two still issue the same statement
once each. Removing that remaining duplication would mean caching
_describe_table_extended() itself, which is a larger change and a different
discussion; happy to open a follow-up issue if you'd like it pursued.

Note on scope of this PR

This started as the get_foreign_keys() fix for #72. While verifying it I found
get_columns() had the identical omission and filed #75. Since both are the same
one-line change to the same file and share one test module, keeping them apart
would have meant two PRs conflicting on all three files — so they are together
here. Happy to split them if you'd rather review them separately.

Completion Criteria

• @reflection.cache added to get_foreign_keys() and get_columns()
• Unit tests for the cache hit, per-table keying, and the no-info_cache path
• Reverting either decorator fails exactly that method's cache test and nothing else
• pytest tests/test_local (offline modules) — 300 passed
• mypy --install-types --non-interactive src — clean (matches the CI check)
• black --check clean on both changed files (_parse.py is flagged only by a
  newer local black than the pinned ^22.3.0, is untouched here, and is
  already flagged on main)
• CHANGELOG entries under a new # Unreleased section
• Commits signed off (DCO)

close #72
close #75