Autotuning - initial infra + a small optimization for pure depth tp

[siddharth9820]

Building the initial infra for autotuning. FC layers can accept a tensor_parallel_dims argument and create a tp decomposition separate from the global tp decomposition.

Noticed that the extra non expert communication takes a considerable amount of time even for pure depth-tp. Note that in this case we aren't actually doing any communication so it's basically the setup costs to launch identity kernels. I simply disabled it.