Difference in encoding CELT media between AMD and Intel CPUs in floating-point mode

Hello,

We recently made note of an issue where encoding CELT media in floating-point mode was giving slightly different results on AMD and Intel CPUs. The difference arose due to the use of _mm_rcp_ps and _mm_rsqrt_ps in op_pvq_search_sse2, see:
https://github.com/xiph/opus/blob/master/celt/x86/vq_sse2.c#L108
https://github.com/xiph/opus/blob/master/celt/x86/vq_sse2.c#L173

These intrinsics/instructions are defined to be approximations, and permitted to have slightly different implementations across CPUs. In order to better produce similar results for offline-generated media across different CPU architectures, it would be good to replace these with _mm_div_ps and _mm_sqrt_ps which have stricter definitions on their behaviour.

The following should be logically similar to the lines above, and harmonize the results across architectures:

For vq_sse2.c#L108:
rcp4 = _mm_mul_ps(_mm_set_ps1((float)(K+.8)), _mm_div_ps(_mm_set_ps1(1.f), sums));
And for vq_sse2.c#L173:
y4 = _mm_div_ps(_mm_set_ps1(1.f), _mm_sqrt_ps(y4));

Start in celt/x86/vq_sse2.c at the two op_pvq_search_sse2 locations linked in the issue, and review how the reciprocal and reciprocal-square-root intrinsics are used. Replace the approximate operations as described, then verify that floating-point CELT encoding produces matching results on AMD and Intel CPUs.