The interactive (current-frame) text query ran with no feedback while the
model processed the frame. Add a progress dialog shown for the duration of
the single-frame query; the all-frames path is unaffected since it runs as a
pipeline job with its own progress UI.

Single-frame text query created every track with the default confidence of
1.0 because TrackStore.add() seeds the confidence pair at 1.0 and the handler
never applied the detection's score. Call setType with det.score so the
model's confidence is reflected instead of always showing 100%.

When a job process exits non-zero (and was not cancelled), scan its captured
output for lines beginning with 'ERROR:' and show them in a dialog. Processes
emit a single 'ERROR: <cause>' line to report the real failure reason (e.g. a
GPU out-of-memory during SAM3 model load), so the user sees the cause without
opening the job log.

Opening the text-query dialog and running a text query previously called
initialize(), which warms the point-segmentation model. Text query uses its
own (often different/heavier) model, so start the service process only
(ensureStarted) and let the text-query model load lazily on first query.

runTextQueryPipeline passed the refiner settings as a flat pipelineParams
object, but runPipeline() only threads pipelineParams.kwiverParams into the
runner's -s flags -- so text_query/replace_existing/detection_threshold were
silently dropped and the refiner ran with its defaults (querying 'object', not
the user's text). Nest them under kwiverParams.