[CIR] Lower __builtin_bswapg (#203618)
C++23 `std::byteswap` lowers every value wider than a single byte
through the type-generic `__builtin_bswapg` builtin, which CIRGen had no
case for, so `std/numerics/bit/byteswap.pass.cpp` hit `errorBuiltinNYI`.
This handles `__builtin_bswapg` the way classic CodeGen does
(`CGBuiltin.cpp`): a bool or single-byte integer is returned unchanged,
and wider values go through `cir.byte_swap`. Unlike the unsigned-only
`__builtin_bswap16/32/64`, the generic builtin also accepts signed
operands, so a signed argument is reinterpreted as unsigned of the same
width before the swap and cast back afterward.
`cir.byte_swap` previously accepted only 16/32/64-bit operands, but
`std::byteswap` instantiates it for `__int128` and wide `_BitInt` too
(the libc++ test reaches `_BitInt(256)`). The operand constraint is
widened to any unsigned integer whose width is a multiple of 16 bits --
which is what `llvm.bswap` requires -- and the existing CIR-to-LLVM
lowering already handles any such width. With the fix,
`byteswap.pass.cpp` passes under `-fclangir`.
