Consider porting `APFloat` support for Power ISA's `long double`

[alexrp]

Background: https://gcc.gnu.org/wiki/Ieee128PowerPC#Background

I do not intend to port this code from LLVM initially. It is known to be wrong in many cases, and I think misleading people into thinking that it's supported correctly is bad. It's notable that GCC just refuses to constant-fold long doubles in this format (breaking with the C standard), while Clang will do it but silently get wrong results. Neither is great, but I think GCC gets it more right than Clang.

Leaving this issue open mainly to gauge interest in this (incomplete) support.