[Feature Request] Add option for no learnable bias in `Conv2d`

[ShreyasKhandekar]

Currently the Conv2d layer in ChAI assumes that there must be a learnable bias for the output. (bias)
Pytorch support the option to turn this off. See https://pytorch.org/docs/stable/generated/torch.nn.Conv2d.html#torch.nn.Conv2d

It would be nice if we can also configure this in ChAI, and also support reading in pytorch dumps of Conv2d layers withoutbias weights when bias=false

[Iainmon]

The bias can be learnable if the convolution layer was separated into two operations: conv + bias. Then the bias would become a learnable parameter. The reason it is no like this is because it is slower for inference. One could inject the bias tensor into the computational graph to get the best of both worlds, but that would be very hacky.

[ShreyasKhandekar]

I've added this by making the bias optional in Conv2D by adding a bais: bool argument to the initializer which can be used to turn off the bias.