We were emitting the wrong instructions for fused multiply-add
operations on LLVM:

  - the instruction name is "llvm.fma.f32" or "llvm.fma.f64", not "fmadd"
  - LLVM does not support other instructions such as "fmsub"; instead
    we implement these by flipping signs of some arguments
  - the instruction is an LLVM intrinsic, which requires handling it
    like a normal function call instead of a machine instruction

Fixes #24223