Compile Graph for Custom Triton Kernels

[vasqu]

Mamba2 relevant issues:

• Mamba2 9 times slower inference time than Mamba1 state-spaces/mamba#355
• triton error while running Mamba2 with slow path state-spaces/mamba#369

Torch relevant issues:

• [Dynamo] torch.cuda.device context manager doesn't work pytorch/pytorch#128059 (seems to occur with 2.3.0)
• torch._dynamo.variables.higher_order_ops error when trying to compile model Dao-AILab/flash-attention#596
• OpenNMT doesn't work with dynamic torch.compile pytorch/pytorch#93468 (comment)

Graph compiling seems to fail due to either torch issues and/or triton-specific unhashables.