Align memory behaviour with the reference Comfy-Org NVIDIA RTX node:
- torch.empty instead of torch.zeros: physical pages fault in as frames
are written, so reserving large video batches no longer forces the
full footprint up front (the OOM source).
- 24 GiB VRAM output ceiling (32 GB card): large batches stay in VRAM
instead of bouncing to the CPU tier.
- Drop the conservative up-front RAM pre-check from the default path;
the kernel decides. 'use_mmap' (now off by default, with a warning in
the help text) is the opt-in last tier of VRAM -> RAM -> disk.
- Collapse per-frame output write to 0 extra buffers in the in-VRAM case
(strided view + in-place copy_); cross-device/dtype still uses one.
- Director recipe + JS UI default follow use_mmap: false.
Tests: 26/26 green; allocator exercised in isolation for both tiers.
Darksidewalker
6785f0dcd6