Fused Ops Support for Lora Dropout

[fabianlim]

This PR addresses #32 to allow Fused Ops to also support lora dropout.

The key strategy is to pass the dropout to the LoRA.apply function, and have it call inside the matmul_lora, and return out the dropped out X to the top-level to be saved for the backward

• LoRA.apply(.., dropout) : dropout is passed into apply.
• matmul_lora(..., dropout) is called with the dropout, in which inside

  X = dropout(X)
  dropout.X = X # saved for retuning to the calling level

• then ctx.save_for_backward(..., dropout.X) is saved for backward (and we will delete it off dropout when it is no more needed)
• then inside backward, we will use dropout.X instead of X.

[fabianlim]

merged to the wrong branch, so I cherry picked to dev on commit 6186ddd4e2dd4c8f53b803bcf032b9fb481ec51e