Qualcomm AI Engine Direct - Enable W2l (#1767)

Summary:
- Based on the reported issue: https://github.com/pytorch/executorch/issues/1342, there are currently some missing operations when lowering W2l to QNN backend.
- Added log_softmax operation
- Added Conv1D. Since QNN does not support Conv1D, this operation is built using unsqueeze to cast it to Conv2D, and then cast the dimension back during output using squeeze.
- Added Squeeze operation
- Changed node_to_wrappers from [torch.fx.Node, PyQnnWrapper.TensorWrapper] to [str, PyQnnWrapper.TensorWrapper]. It will use node.name as key instead of node because it allows more flexibility in creating extra tensors and nodes for QNN op builders.
- Currently, no scripts are added for W2l as it encountered "No audio backend is available" when using torchaudio. Another PR will be pushed in the future when this this issue can be resolved.

Pull Request resolved: https://github.com/pytorch/executorch/pull/1767

Reviewed By: larryliu0820

Differential Revision: D53507975

Pulled By: cccclai

fbshipit-source-id: fa1e58c9c05858efe7a7f11340a6dc265c7a8390