Fix partitioning for constants (#3991)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/3991

After we run the constant prop pass the weights/biases might become constants in the exported program which we're not looking for currently when partitioning the graph. This diff fixes that and also ensures that the update is properly done in `unsqueeze_kernel_weights`.

Reviewed By: mcr229

Differential Revision: D58399313

fbshipit-source-id: ce0e747759b287e3a70bbb7ccb95a26d11e949d7