[Runtime] Fix overload resolution when targeting Haswell.

[allevato]

I hit this issue when I tried targeting Haswell instead of Sandy/Ivy Bridge with our internal build of the toolchain:

Float16Support.cpp:165:10: error: call to 'fromEncoding' is ambiguous
  165 |   return fromEncoding(__gnu_f2h_ieee(f));
      |          ^~~~~~~~~~~~
Float16Support.cpp:45:14: note: candidate function
   45 | static float fromEncoding(unsigned int e) {
      |              ^
Float16Support.cpp:59:17: note: candidate function
   59 | static _Float16 fromEncoding(unsigned short s) {
      |                 ^

I thought it was an issue with our setup, but it repros in isolation on Godbolt: https://godbolt.org/z/9hTqrzPvr

Switching the __F16C__-guarded definitions from short to unsigned short to match the other signatures appears to resolve the issue in our builds.

[allevato]

@stephentyrone Could you double-check this for me? This isn't a space I'm too familiar with, so I wasn't sure if they were specifically short for a reason.

[stephentyrone]

They were short because that's how these were originally defined in compiler-rt ages ago, but they're uint16_t now, so this ought to be fine.

[allevato]

@swift-ci please test and merge

[allevato]

@swift-ci please test