[iris] Simplify replay system: tests-only, drop SQL traces, flatten goldens

Address review feedback:

- Replay code lives under ``lib/iris/tests/cluster/controller/replay/``
  instead of ``lib/iris/src/iris/cluster/controller/replay/``. It's
  test infrastructure; belongs with the tests.
- Drop the SQLite ``set_trace_callback`` hook and the ``sql_trace``
  module. The trace was dominated by 500+ migration DDL lines and the
  DB-state dump already catches every behavioral drift we care about.
  Reverts the ~10-line ``db.py`` addition.
- Flatten golden tree: ``golden/<scenario>/db.json`` -> ``golden/<scenario>.json``.
  ``sql.txt`` goldens deleted.
- Merge ``events.py`` and ``dispatcher.py`` into a single module.
- Move ``frozen_clock`` out of every scenario into a top-level context
  manager entered once by the test runner, after ``ControllerDB``
  construction.
- Delete the ``run.py`` CLI and ``__main__.py``. Running
  ``pytest lib/iris/tests/cluster/controller/replay/`` is the full
  workflow; a CLI can be added back if marin-checkpoint-seeded runs
  are wanted.

Also from the PR-B review:
- ``register_or_refresh_worker`` docstring updated: the attr-cache
  update and audit line are scheduled as ``cur.on_commit`` hooks,
  not returned to callers.
- ``service.launch_job`` closes the submit-with-replace TOCTOU by
  running the existing-job check and the conditional cleanup in a
  single transaction. ``submit_job`` still opens its own tx
  downstream; a concurrent submitter between the two txs collides on
  the jobs PK (legitimate error, not corruption).

Tests: 815 passed on the controller suite; 13 replay goldens match
exactly. pyrefly clean in the controller tree.

Author: rjpower

lib/iris/src/iris/cluster/controller/db.py

@@ -294,14 +294,6 @@ class ControllerDB:
     AUTH_DB_FILENAME = "auth.sqlite3"
     PROFILES_DB_FILENAME = "profiles.sqlite3"
 
-    # Class-level SQL trace hook: when set, every connection (writer + read pool)
-    # opened via _configure() registers this callback via
-    # ``conn.set_trace_callback``. Exposed for the replay-system tests
-    # (``iris.cluster.controller.replay``); production never sets this. Must be
-    # set BEFORE ``ControllerDB(...)`` is constructed so the writer connection
-    # picks it up — connections opened earlier will not retroactively trace.
-    _trace_callback: Callable[[str], None] | None = None
-
     def __init__(self, db_dir: Path):
         import time
 
@@ -423,8 +415,6 @@ def _configure(conn: sqlite3.Connection) -> None:
         conn.execute("PRAGMA synchronous = NORMAL")
         conn.execute("PRAGMA busy_timeout = 5000")
         conn.execute("PRAGMA foreign_keys = ON")
-        if (cb := ControllerDB._trace_callback) is not None:
-            conn.set_trace_callback(cb)
 
     def optimize(self) -> None:
         """Run PRAGMA optimize to refresh statistics for tables with stale data.