Decoupled PerChannel/PerTensor quantization

[Giuseppe5]

To have a decouple per channel pre-scale, and per tensor post scale:

class Int8WeightNormL2PerChannelPerTensorFixedPoint(Int8WeightNormL2PerChannelFixedPoint):
    scaling_per_output_type = ScalingPerOutputType.TENSOR

It currently works for all quantizers that inherit from WeightNormPerChannelFloatDecoupled

It can be extended to other decoupled quantization quantizers if needed

[nickfraser]

Single question about test organisation (specifically, fixtures), otherwise LGTM!

[nickfraser]

This is a code style question, any reason why we don't use the standard pytest way of sharing fixtures across multiple files (i.e., conftest.py)?

[Giuseppe5]

We have a lot of duplication for various reasons and in general tests would need to be restructured a bit.
All this to say that I wouldn't be sure at his point of the cleanest way to do it myself, but if you have a suggestion, all ears.