Cancel does not stop in-flight GPU inference

[taskmasterpeace]

Severity: High (confirms a known issue, with precise location)

backend/handlers/video_generation_handler.py (~L269-285; A2V path ~L538-557).

pipeline_state.pipeline.generate(...) runs synchronously to completion. is_generation_cancelled() is checked only before (~L205) and after (~L282) inference — never during. cancel_generation() merely swaps state to GenerationCancelled; the in-flight generate() (the multi-minute/hour part) has no cancellation token and runs to the end. Same pattern in the API pollers.

Fix

Thread a cancellation callback / threading.Event into the pipeline's denoising loop so it can bail between diffusion steps.