[QuanitztionModifier] initialize per channel scales and zps to correct shape

[bfineran]

fixes bug found by @anmarques
pytorch does not initialize channel wise qat scales and zero points to their final shape, just to (1,). this adds a patch to expand them to their expected shape

test_plan:
@anmarques to fully reproduce and verify

existing unit tests pass

manual inspection of initialized scales and zero points:

from sparseml.pytorch.sparsification import QuantizationModifier
from sparseml.pytorch.models import resnet50

r = resnet50()
QuantizationModifier(scheme=dict(weights=dict(strategy="channel"))).apply(r)
weight_quants = [(n, mod) for n, mod in r.named_modules() if "weight_fake_quant" in n]

[anmarques]

Tested on Llama2