feat: --omit-schema flag to drop the sole schema's qualifier from output

[Apakottur]

What

Adds --omit-schema NAME to pgmig generate (and omit_schema= to pgmig.generate): omits that schema's qualifier from the emitted SQL, for a readable diff when a database uses a single schema (typically public). Guarded by validation that NAME is the only user schema in both databases — otherwise a clean PgmigError — so an unqualified name can never be ambiguous.

Behavior

With --omit-schema public (single-schema databases):

source	target	emitted SQL
absent	person table	CREATE TABLE "person" ("id" integer);
absent	index on person	CREATE INDEX person_id_idx ON person USING btree (id);
absent	FK to team	ALTER TABLE "person" ADD CONSTRAINT "person_team_fk" FOREIGN KEY (team_id) REFERENCES team(id);
absent	function	CREATE OR REPLACE FUNCTION bump(n integer) ...

Without the flag, output is unchanged (fully qualified, as today). With the flag but a second user schema present (or the name wrong): omit_schema 'public' requires the source database to contain exactly that one user schema, but found: [...].

Applies to every emitted object path: tables (incl. OWNER TO), columns, indexes, constraints, triggers, functions, sequences, enums, views, materialized views, domains and composite types.

Notes

The rendering policy is a field on the diff context introduced in #80: context_scope(..., omit_schema=...) carries it for the duration of one generate call, and the rendering helpers in _sql.py read context.omit_schema (lenient — None outside any diff scope, so they stay usable as plain functions). Emit sites call schema_qualified(schema, *rest) with no parameter threading through the diff helpers.

No single mechanism covers all emitted SQL, so this works in three legs:

• pgmig-built paths go through schema_qualified, which drops the matching schema segment.
• Server-generated definitions (pg_get_indexdef/functiondef/triggerdef/constraintdef/viewdef) are introspected with the schema on the search_path, so deparse drops the qualifiers it resolves semantically (FK REFERENCES targets, trigger EXECUTE FUNCTION, view bodies) — without ever touching string literals. This intentionally trades the empty-search_path portability property from fix: introspect with empty search_path for deterministic, portable SQL #58 when the flag is used; the default (omit_schema=None) keeps the empty search_path.
• Spots deparse always qualifies regardless of search_path — the index/trigger ON schema.table clause and a routine's own header name — are edited textually by anchored helpers (strip_on_clause_qualifier, strip_routine_name_qualifier) that try each identifier in its quoted and unquoted deparse form. If no form matches, the definition is returned unchanged (valid SQL, just qualified) rather than risking a wrong edit.

Deliberately untouched: CREATE SCHEMA / COMMENT ON SCHEMA and the extension SCHEMA x clause name the schema as their subject, not as a qualifier. Error diagnostics also honor the omission (unlike an earlier revision of this PR): validation guarantees a sole user schema, so an unqualified diagnostic stays unambiguous.

Existing tests are untouched (default path unchanged); new coverage: unit tests for schema_qualified, both strip helpers (quoted/unquoted/fallback forms) and the context-scope reset (incl. on exception), plus integration tests proving no public token survives anywhere in the output across every supported object kind — and that the migration still converges on apply.