Make _FloatN and _BitInt types as substitution candidates in mangling

[zsrkmyn]

As discussed in https://reviews.llvm.org/D140359, encodings of _FloatN and _BitInt(N) are relatively long, and thus should be treated as substitution candidates to shorten mangled names.

[rjmccall]

Substantively, this seems reasonable to me.

[zsrkmyn]

May I know if there's any concern on this? :-)

[zsrkmyn]

@rjmccall, can we get this checked in? Or is there anything else I should do? thanks!

[rjmccall]

Sorry for losing track of this. Three comments:

1. Given that we've also added manglings for _Float16x and std::bfloat16_t, I think we need to understand the exact set of types we want to allow to serve as substitution candidates.
2. These extensions have been around for awhile, so I'd like to understand what existing practice is here and make sure that we have agreement between vendors implementing it.
3. It'd be nice if we had a more precise way of specifying this list, maybe by splitting up the grammar a bit so that the substitutable builtin types were a different production from the unsubstitutable ones.

[al13n321]

FWIW, current clang substitutes _BitInt(N), but not _FloatN:

% llvm-cxxfilt _Z1fDB8_1xS_ 
f(_BitInt(8), x, _BitInt(8))

% llvm-cxxfilt _Z1fDF32_1xS_
f(_Float32, x, x)

% llvm-cxxfilt --version   
llvm-cxxfilt
LLVM (http://llvm.org/):
  LLVM version 21.1.5
  Optimized build.