Fixes #77658

This improves codegen mostly for unsigned to floating types but catches a few other redundant conversions.

• Adds support for using AVX-512 vcvtusi2s[sd] for uint -> float/double (only ulong was handled previously) on both x64 and x86.

  -       mov      eax, edx
          vxorps   xmm0, xmm0, xmm0
  -       vcvtsi2sd xmm0, xmm0, rax
  -       vcvtsd2ss xmm0, xmm0, xmm0
  +       vcvtusi2ss xmm0, edx

  -       mov      eax, dword ptr [rbp-0x04]
  -       mov      eax, eax
          vxorps   xmm0, xmm0, xmm0
  -       vcvtsi2sd xmm0, xmm0, rax
  +       vcvtusi2sd xmm0, dword ptr [rbp-0x04]

• Improves codegen for uint -> float conversions on x64 without AVX-512, removing the intermediate conversion to double.

          mov      eax, edi
          xorps    xmm0, xmm0
  -       cvtsi2sd xmm0, rax
  -       cvtsd2ss xmm0, xmm0
  +       cvtsi2ss xmm0, rax

• Adds support for direct ulong -> float cast to the x64 SSE2 fallback, resolving a difference in behavior between hardware with AVX-512 vs without, and saving an extra cvtsd2ss instruction.

          xorps    xmm0, xmm0
          mov      rax, rdi
          shr      rax, 1
          mov      rsi, edi
          and      rsi, 1
          or       rsi, rax
          test     rdi, rdi
          cmovns   rsi, rdi
  -       cvtsi2sd xmm0, rsi
  +       cvtsi2ss xmm0, rsi
          jns      SHORT G_M37561_IG56
  -       addsd    xmm0, xmm0
  +       addss    xmm0, xmm0
   G_M37561_IG56:
  -       cvtsd2ss xmm0, xmm0

• Removes some redundant float -> double -> float casts.

  -       vcvtss2sd xmm1, xmm1, xmm1
  -       vcvtsd2ss xmm1, xmm1, xmm1
          vbroadcastss xmm1, xmm1

SPMI Diffs

The only code size regressions are the insertion of xorps to clear the upper elements of the target reg for the AVX-512 unsigned conversion instructions. These were previously omitted but should have been there since the unsigned conversions have the same behavior as the signed (i.e. preserving/copying upper elements) and are subject to the same false dependency penalties.

GCC emits the xorps for all conversions; Clang skips it for all conversions in simple examples but may emit it in more complex scenarios.
https://godbolt.org/z/6aY7fdE3d