fix symmetric zero points for unit8 quantization

[bfineran]

to accomodate custom quantization ranges, in #542 we added a pass to propagate quantization range information from FakeQuantization objects to their Observer objects (this functionality is missing due to a small pytorch bug).

An unintended side effect, is that now, for default symmetric uint8 quantization (range [0, 255]), the range is now considered by torch to be custom, thus calculated as 255 - 0 // 2 = 127 instead of the desired 128 which was hardcoded as the default before.

This changes addresses this bug by skipping over the propagation pass for objects with default symmetric quantization ranges.

Testing can be done using the second input of QATMatMul from our BERT integration which should have a symmetric zero point of 128. Prior to this patch, the example shows it to be 127.

from transformers.models.bert.modeling_bert import QATMatMul
from sparseml.pytorch.sparsification import QuantizationModifier
import torch

dummy_sequential = torch.nn.Sequential(QATMatMul())  # give QATMatMul a layer to be wrapped
QuantizationModifier().apply(dummy_sequential)
qat_matmul = dummy_sequential[0]
_ = qat_matmul(torch.randn(10, 10), torch.randn(10, 10))
# print affected zero point, 127 before fix, 128 after
print(qat_matmul.input_quant_stubs[1].activation_post_process.zero_point)