Runs out of memory on 128Gb Mac after processing compete due to buffer management issue

[RDouglasSharp]

1. ComfyUI output conversion — triple copy spike (video_upscaler.py L516-530)

sample = ctx['final_video']          # Reference, no copy yet
sample = sample.cpu()                # NEW copy on CPU (fp16). GPU copy still alive via ctx.
sample = sample.to(torch.float32)    # NEW copy on CPU (fp32). Old CPU fp16 still alive.
# PEAK: GPU fp16 + CPU fp16 + CPU fp32 = ~3x video size simultaneously

Please fix this triple buffer issue.  My mac has completed upscaling, and then the memory blows up due to 3 copies being made sending the result back to ComfyUI!