[ingress][mlir-gen] Merge gpu payload generation methods to mlir-gen

[tkarna]

Merges "gpu" payload generation methods to mlir-gen.

• Moved ingress/gpu/* files to ingress/mlir_gen/ and refactored code to use shared methods.
• Removed value/tensortype dualism from mlir_gen op generation functions.
• Gpu matmul payload generator uses mlp payload generator directly.
• Changed mlp payload function argument order to: output, input, weights, [biases].

[fschlimb]

Maybe assert that truncf/extf are valid, or rename the func so that it doesn't sound as if it can do arbitrary conversion.

[tkarna]

Updated

[fschlimb]

I wonder if we can come up with a less fragile way of relying on the generated function name. The payload function name is parametrizable, this here (as well as other util funcs) are hard-coded here and "guessed" on usage. Not a show-stopper, but the current approach is not very nice (same in the mlp-mpi example and probably others).

[tkarna]

Yes, I agree. At least when with respect to alloc/dealloc functions I hope a MemoryManager abstraction will fix this. The manager creates the helper funcs and also calls them, so the naming scheme is defined in one single place.

[adam-smnk]

Good reorg, perfectly fits the mlir-gen module.