Support arbitrary data bits

[dbericat]

Is your feature request related to a problem? Please describe.

• Buffers are not floating point buffers (integers or less than integers 8-4-4 or even 4-2-2)
• Limitations in how data flows around in GStreamer (video format, can’t pass floating points numbers, no point in transforming before inference)
• Transforms -> inference -> transforms flow can’t be done. Hast to be done as G streamer data types.

Describe the solution you'd like

• Goal: Read into CUDA tensors
• DS fix: parameters on inference that handles offset

Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.

Additional context

• Need use cases: current models from KCL had 1 channel input. Had to retrain to have 3 channels
  • Select from 2-3 # of channels down to 1
• How do we insert pre-transforms
  • Can’t do it in a separate component.
  • Enhance DS inference server