[llvm] [AMDGPU] Add test for v_fmamk_f16/v_fmaak_f16 in real-true16. NFC (PR #173307)

[Stanislav Mekhanoshin via llvm-commits]

rampitec wrote:

> The design here is supposed to be that we do not shrink vop3 true16 to shorter encodings like vop2 etc pre-RA. So RA only sees V_FMAC_F16_t16_e64. Then in SIShrinkInstructions post-RA, if the instruction can be legally shrunk to FMA* with the given RA we do it. This provides maximum freedom to the RA, instead of restricting it to put those values in only the first 128 registers (a tradeoff that pessimizes encoding size a bit). Do you want to deviate from that design?

So, yes I want to deviate from that design. We will still do allocation as before, but I also want lo128 to be allocatable for the `fmaak` and `fmamk` VOP2 only instructions and be able to produce these not just post-RA.

https://github.com/llvm/llvm-project/pull/173307