• Fixed integer overflows which could cause out-of-bounds reads or writes
  instead of an error if a very large byte offset was passed to
  oidnReadBuffer, oidnWriteBuffer or oidnSetFilterImage, or a very large
  pixel or row stride was passed to oidnSetFilterImage or
  oidnSetSharedFilterImage
• Fixed oidnSetSharedFilterImage accepting a byte offset large enough to
  make the address of the image overflow instead of returning an error
• Fixed oidnSetFilterImage and oidnSetSharedFilterImage truncating the
  width and height of the image to 32 bits, which could silently denoise a
  much smaller image instead of returning an error
• Fixed a crash if a null array of command queues was passed to
  oidnNewMetalDevice, which now returns an error instead
• Fixed leaking the imported memory object on CUDA and HIP devices if
  oidnNewSharedBufferFromFD or oidnNewSharedBufferFromWin32Handle failed
  after the handle had already been imported
• Fixed oidnReadBuffer and oidnWriteBuffer failing with an out-of-memory
  error on Metal devices if the specified byte size was 0
• Fixed a signed integer overflow when denoising very large single-channel
  images, close to the maximum supported number of pixels
• Fixed out-of-bounds reads and crashes caused by insufficient validation of
  the weights blob set with oidnSetSharedFilterData, which could occur if
  the blob was corrupted or malicious
• Fixed a crash (division by zero) if the weights blob set with
  oidnSetSharedFilterData contained a tensor with a zero dimension
• Fixed changes to other filter parameters being discarded when setting an
  empty data parameter with oidnSetSharedFilterData, which could cause the
  filter to be executed with an outdated model, corrupting the output image or
  crashing
• Fixed crashes and incorrect output if committing a filter failed due to
  running out of memory, which could also break the other filters of the
  device. Executing such a filter now returns an error instead, and the filter
  can be recovered by committing it again