…ware-fixes

Co-authored-by: Copilot <copilot@github.com>

Exclude examples/inference/franka.py from mypy to resolve a duplicate
"franka" module clash with examples/teleop/franka.py (whose franka.py
is the one example/teleop/quest_align_frame.py actually imports from).

Make LimitedAbsoluteAction._get_current()'s return annotation honest
(np.ndarray | common.Pose) since it returns a Pose in cartesian control
modes, and add a cast at the one call site that lacked one. Add
rule-specific type: ignore / noqa comments for the remaining cases
where the underlying stubs don't match runtime behavior (FR3-only
robot methods, duckdb's loosely-typed schema introspection, Optional
fields that are known to be set, and the lerobot torch imports).

StorageWrapper._flush() inferred and permanently cached its pyarrow
schema from whatever rows happened to be in the first flush. Since
"action" is logged one step behind (frame 0 has no prior action), an
unlucky first batch consisting only of that null action got pyarrow's
null type pinned for the field forever, so every later batch with a
real action value failed with "Invalid null value". Defer flushing
until the buffered rows yield a schema with no null-typed fields
instead of caching a too-eager guess.

Separately, close() called _flush() and only sent the writer thread's
stop sentinel afterwards, so any exception during that flush (such as
the one above) left the background ThreadPoolExecutor thread blocked
on an empty queue forever, hanging the whole process at interpreter
exit. Wrap the flush in try/finally so the sentinel is always sent.

Also add the include_rotation field quest.py now reads to the
QuestOperator test config double, which test_consume_action_swaps_
gripper_with_controller was missing.