nir: Model integer multiply-add
Both AGX and NVIDIA have this instruction, so let's add it to NIR without a vendored suffix. We don't add rules to fuse imad from imul+iadd at this point, since that needs more thought and can be done in a backend algebraic pass as long as the opcode itself is introduced.
Signed-off-by: Alyssa Rosenzweig alyssa@rosenzweig.io