Quantization refactor

[anmarques]

Changes in quantization:

1. Created flag to remove Q/DQ from output activations for conv layers. Set to True by default.
2. Changed default to True for flag to remove Q/DQ from output activations for linear layers.
3. Created flag to change number of bits for weight activation (similar to the already existing flag for activation bits).
4. Created flag to avoid quantization of BatchNorm layers when fusing is not used. Set to True by default.
5. Created flag to enable quantization compatible with tensorrt. Set to False by default.
   • Sets activation quantization to symmetric and UINT8.
   • Sets default fusing function to 'no_fuse'.
6. Created a wrapper module (BNWrapper) that allows one to freeze BatchNorm statistics when BN layers are not fused. PyTorch default only works when BN layers are fused.
7. Crated a wrapper module (_AddReLU) for the FloatFunctional class that performs the Add + ReLU operations in ResNet. This wrapper enables the module to be wrapped by a QATWrapper when quantized, and is set to quantize only the first input.

[bfineran]

merged in fixes for testing and additional tests from #685

contributor

[spacemanidol]

look good