Update on "[reland][quant][fix] Add bias once in conv_fused (#48593)"

Summary:
Previously _conv_forward will add self.bias to the result, so bias is added twice in qat ConvBn module
this PR added a bias argument to _conv_forward and _conv_forward is called with zero bias
in ConvBn module

fixes: #48514

Test Plan: Imported from OSS

Reviewed By: raghuramank100

Differential Revision: [D25249175](https://our.internmc.facebook.com/intern/diff/D25249175)

[ghstack-poisoned]

torch/nn/intrinsic/qat/modules/conv_fused.py

@@ -93,7 +93,8 @@ def _forward(self, input):
         bias_shape = [1] * len(self.weight.shape)
         bias_shape[1] = -1
         scaled_weight = self.weight_fake_quant(self.weight * scale_factor.reshape(weight_shape))
-        # this does not include the conv bias
+        # using zero bias here since the bias for original conv
+        # will be added later
         if self.bias:
             zero_bias = torch.zeros_like(self.bias)
         else: