Support Linear Fused Batchnorm

[mcr229]

🚀 The feature, motivation and pitch

Currently XNNPACK only supports batchnorm when it succeeds a convolution. This is because we can fuse it with the preceeding convolution operation. However we can also do this for Linear layers, by fusing batchnorm with the previous linear. Take a look at the code for fusing with conv, and try to add a pass to do the same for linear:

https://github.com/pytorch/executorch/blob/main/backends/xnnpack/_passes/fuse_batch_norm_with_conv.py

Alternatives

No response

Additional context

No response

RFC (Optional)

No response

[keyprocedure]

I’d like to take this on

[mcr229]

assigning it to you!

[keyprocedure]

There's quite a bit of duplicated logic between fuse_batch_norm_with_linear.py and fuse_batch_norm_with_conv.py.
Would it be better to refactor it in this PR, or follow up with another PR after the current one lands?