Study type packing (VNNI/BFDOT/BFMMLA/etc) as a single operation

[rengolin]

Today we're working on a type packing for VNNI with the operation tpp.vnni_pack. But this isn't the only type of packing we may want, and they're all very similar, so we could probably come up with a single op (say tpp.type_pack) with generic arguments that converts a standard packed tensor into a type packed tensor, given some shapes.

We then map the existing types (VNNI/BFDOT/BFMMLA/etc) into this op and lower them as library calls when they match a recognised shape.

We shouldn't create this op if we don't know that the type packing makes sense (ex. by #561 calls), so we shouldn't need to lower this op to loop unless requested by the pipeline (tpp-to-loops).

@KavithaTipturMadhu

References:

• VNNI: https://www.intel.com/content/www/us/en/developer/articles/guide/deep-learning-with-avx512-and-dl-boost.html
• BFDOT: https://developer.arm.com/documentation/ddi0602/2021-06/SVE-Instructions/BFDOT--vectors---BFloat16-floating-point-dot-product-
• BFMMLA: https://developer.arm.com/documentation/ddi0602/2023-03/SIMD-FP-Instructions/BFMMLA--BFloat16-floating-point-matrix-multiply-accumulate-into-2x2-matrix-?lang=en