9.40

• Speed up --vpp-kfm.
• Move single-frame, mask, and multi-frame pre/post-processing in --vpp-onnx to CUDA.
• Fix --vpp-onnx reading a stale frame before the inference result is updated under GPU load.
• Fix a race in raw output (--codec raw) where a GPU buffer could be reused during transfer.
• Add a PAL/SECAM VHS restoration filter --vpp-vhs-restore. ( rigaya/HWEnc-onnx-models#1 )
  Models are added to the HWEnc-onnx-models deinterlace archive.