Introduce bin_map_pk2

[ruanjm]

Make operations on tile can use pk instruction. However, this design is not decent to me. Feel free to share your idea and deprecate this design.

Besides, I did try to use if constexpr to check whether a class or structure has member function op_pk2(). However, this design doesn't work with latest hipcc... Hipcc still tries to compile the false path even if it knows the result of condiction at compile time.

[willhu-jpg]

Ideally, we can have a if constexpr to check whether a class or structure has the op_pk2() member function.

Alternatively, we should just create a mul_pk struct and add a contexpr bool hasPacked; field to every operator struct. That way we can have a single bin_map function.

bin_map functions that operate over a single register should throw a compile-time error if it's used with a packed instruction. Otherwise, the other bin_map functions would just skip over every other register when the operator has hasPacked set to true.

The developer would now have the flexibility of choosing between using a packed or non-packed instruction.

[willhu-jpg]

I think instead of hk::bin_map_pk2<0, 0, hkm::mul_vgpr>(oaccu, oaccu, rescale);, we can make things a bit prettier in the library and do something like hk::mul_vgpr_pk2<0, 0>(oaccu, oaccu, rescale);.