fix(#1438): preserve json flag for MariaDB longtext-aliased columns

[dimitri-yatsenko]

Summary

Fixes #1438. MariaDB stores json columns as longtext and reports them back through information_schema as longtext, so DataJoint's DB-type-based JSON detection in Heading._init_from_database() left attr["json"] as False. That silently broke insert (json.dumps skipped → dict reaches pymysql → TypeError) and fetch (json.loads skipped → caller gets the raw string).

The :json: comment marker is already written at declare-time (JSON is in CORE_TYPE_NAMES / SPECIAL_TYPES) and already parsed back into attr["original_type"] on read. Recover attr["json"] from that marker in the same block where UUID is already recovered.

Behavior by backend

• MySQL: DB reports json, regex matches, attr["json"] = True from the existing line. The new branch reasserts the same value — no behavior change.
• PostgreSQL: DB reports json/jsonb, regex matches both. Same as MySQL — no-op.
• MariaDB: DB reports longtext, regex misses, attr["json"] was False. The :json: comment marker is intact, so category == "JSON" and the new branch flips it to True. Insert and fetch now round-trip correctly.

Out of scope

• Legacy MariaDB tables created without the :json: comment marker (e.g. created by a non-DataJoint tool) will still fail. Adding a heuristic for bare longtext columns would be unsound.
• MariaDB CI coverage. MariaDB remains unofficially supported and will not be added to this repo's CI matrix. Downstream projects depending on MariaDB own MariaDB testing in their own pipelines.

Test plan

• Round-trip the issue reproduction against a MariaDB container (insert dict, fetch back dict).
• pytest tests/integration/test_json.py against MySQL — confirm no regression.
• pytest tests/integration/test_json.py against PostgreSQL — confirm no regression.
• Spot-check heading.attributes["metadata"].json is True and .type == "longtext" on MariaDB after the insert succeeds.