* common.h: Centralize RYU_32_BIT_PLATFORM. This indicates whether we
should avoid uint64_t operations for performance, and is activated by
the MSVC platform macros _M_IX86 and _M_ARM (also defined by Clang on
Windows). This centralization will make it easier to detect additional
platforms as 32-bit. Note that if a 32-bit platform isn't detected
here, the consequences are limited to performance, as Ryu will use
correct-but-slower uint64_t operations.

* d2s_intrinsics.h: Include common.h for RYU_32_BIT_PLATFORM.

* Fix comment, as the shift value can be 49. (Example:
`r = 0xba6fa7161a4d6e0c` with `ieeeMantissa = 0x000fa7161a4d6e0c` and
`ieeeExponent = 0x000003a6`.) I've experimentally observed that the
upper bound might be 57 instead of 58, but I can't prove that
theoretically.

* Optimize shiftright128() for 32-bit platforms. If we can assume
`dist >= 32` (which the optimizer can't see), then we can grab lo's
most significant 32 bits with `(uint32_t)(lo >> 32)` (32-bit compilers
can do this efficiently), and then perform a 32-bit shift for the
remaining `(dist - 32)`. Additionally, `uint64_t | uint32_t` is
less work. I observe that this improves Clang 7 RC1 perf by 2.8%
(80.382 ns improved to 78.193 ns) which seems nice for such a small,
localized change. I also observe that this degrades MSVC perf by 0.6%
(106.155 ns degraded to 106.759 ns), but (1) that's very small and
(2) Clang is an existence proof that MSVC can do better (tracked by
VSO#634761).

* Extend `#if defined(_M_IX86)` to `#ifdef RYU_32_BIT_PLATFORM` and
update comments accordingly. This extends the
division-by-64-bit-constant workaround to ARM32 (for which LLVM#37932
was originally filed).

* f2s.c: This already includes common.h, so we can use
RYU_32_BIT_PLATFORM with no behavioral changes.