[Dynamo] Ignore no-op tensor assignment (#106092)
Ignore no-op `self.attr = self.attr` on NN Modules when attr is a Tensor attribute.
This comes from a [llama pattern](https://github.com/pytorch/benchmark/blob/main/torchbenchmark/models/llama/model.py#L121-L122). Normally, when a set attr occurs on an nn module we turn it into an `UnspecializedNNModuleVariable` which prevents static buffers and parameters. In subsequent pr i will add support for cudagraph mutation of buffers/params, which with this pr takes llama 1.6x -> 4.4x in inference
Pull Request resolved: https://github.com/pytorch/pytorch/pull/106092
Approved by: https://github.com/yanboliang
