BatchNorm + Convolution fusion in `prepare_pt2e` removal

[skywall]

🚀 The feature, motivation and pitch

There is BatchNorm + Convolution fusion happening in prepare_pt2e function call. I don't think it's good practice to introduce optimizations randomly into the pipeline (prepare_pt2e function call) because this might be a side effect for the user of an API.

I'm proposing removal of this optimization and letting user peform this optimization on his own, before quantization.

Follow-up from: pytorch/executorch#10579 (comment)
cc @jerryzh168 @jianyuh @raghuramank100 @jamesr66a @vkuzo @jgong5 @Xia-Weiwen @leslie-fang-intel @msaroufim @digantdesai

Alternatives

No response

Additional context

No response

[jerryzh168]

I think this is reasonable, do you want to start a proposal for APIs, and how to guarantee BC

also I'll move this to torchao since we are moving pt2e flow to torchao as well.