This version varies slightly from the current reference implementation in TFLite (original):
1.) All references to tensor_utils:: namespace are dropped due to the build incompatibility and size the include brings in (just like the rest of this port for float/hybrid-quant).
2.) Scratch tensors are re-worked into variable tensors. This is a temporary workaround until memory planning lands .
3.) An additional Tensor is required to provide pre-calculated scale values. These calculations are very expensive on low power device.
PiperOrigin-RevId: 286278125
Change-Id: Ibbadb2f38a6c25b5550b4fced5b32c7a5b9420df
