Python kernel interface should be reworked to accept CBs rather than tensors

[TT-billteng]

Right now, the python kernel interface accepts tensor shapes as kernel arguments as this code was mainly for discovery / prototyping many aspects of MLIR development. I'd like to give users an experience that's closer to traditional metal kernel development in C++ today and that means using CBs as primary method of kernel I/O.

Are there other arguments we should pass via kernel function args? TBD