[clang] Add `__builtin_fmaximum` which lowers to `llvm.maximum`

This is a feature request.

Currently, there exists a `__builtin_fmax` and a `__builtin_fmaximum_num` to implement the corresponding `` functions, but there exists no `__builtin_fmaximum`.

Such a portable builtin would be quite useful:
- it implements `fmaximum` / `std::fmaximum` directly
- it already has a portable `llvm.maximum` intrinsic backing it
- on WASM, it lowers to a single `f32.max` or `f64.max` instruction
- on ARMv8, I think it lowers directly to `fmax`, though I'd need to double-check that
- it implements the ISO/IEC 60559 "maximum" operation

Start by tracing the existing __builtin_fmax and __builtin_fmaximum_num implementations and their lowering to llvm.maximum. Add the missing __builtin_fmaximum behavior and verify that it lowers to the llvm.maximum intrinsic, including the f32 and f64 WebAssembly cases.

Written by the indexing model from the issue text.