[llvm] [AMDGPU][True16] Support V_CEIL_F16. (PR #73108)

[Ivan Kosarev via llvm-commits]

kosarev wrote:

Parsing True16 registers here slightly differs from the downstream implementation where we use an `IsHigh` flag and rely on that only 16-bit registers can be high or low and that all 16-bit registers are either high or low. Using the enumeration is hoped to make the parsing code a bit more explicit.

Tests are pre-committed, but the commit is part of this very pull request. Would that be okay?

For this pilot change I'm going to prepare a corresponding downstream patch; further upstream changes of the kind should be trivially mergeable, I expect.

https://github.com/llvm/llvm-project/pull/73108