Enable cudagraph for `pre-mlp to attn` memcpy

[GindaChen]

Description

#119 introduces the general CUDA Graph construct that captures the MLP portion of the model forwarding. There is one left over: there are a few small memcpy kernels that is not captured by the CUDA Graph, and the current implementation will have invalid memory access, likely from inside of the kernel.

This commit document the issue: (2ed558d)

[GindaChen]

cc @wu-qing-157