Optimize uint64_reverse_bits, rand_int, and count_lt/gt

[mulugetam]

Replace three utility functions with faster algorithms. All changes are semantically equivalent (same outputs for all inputs).

Changes:

faiss/utils/byteswap.h (new)
Portable byteswap32/byteswap64 wrappers (__builtin_bswap* on GCC/Clang, _byteswap_* on MSVC), following the pattern of popcount.h.

faiss/utils/hamming.cpp
uint64_reverse_bits: replace a 64-iteration serial bit-shift loop with byteswap64 + 3 parallel nibble/crumb/bit mask-shift pairs. The loop has 64 data-dependent operations; the new version has 7 independent operations and runs in ~2 cycles.

faiss/utils/random.cpp
rand_int(int max) in RandomGenerator and SplitMix64RandomGenerator: replace rng() % max with Lemire's fast-range (uint64_t(rng()) * uint64_t(max)) >> 32. Eliminates integer division and modulo bias.

faiss/utils/utils.cpp
count_lt / count_gt: replace O(p) linear scan with O(log n) std::lower_bound on the already-sorted input arrays. lower_bound compiles to branchless cmov instructions, avoiding the branch mispredictions and O(p) scan latency that hurt the linear scan at high hit rates. In practice n equals k, which defaults to 1024 in range_search_gpu, putting every call in a regime where binary search wins at all hit rates.

benchmark: https://gist.github.com/mulugetam/524854f16b20f236e645f2cb6da24fc0

Results (measured on Sapphire Rapids):

reverse_bits:

| Function                        | Before    | After    | Speedup |
|---------------------------------|-----------|----------|---------|
| uint64_reverse_bits             | 34.2 ns   | 2.99 ns  |  11.4x  |

Lemire's fast range:

| Function                        | Before    | After    | Speedup |
|---------------------------------|-----------|----------|---------|
| rand_int(max)                   |  3.96 ns  | 2.35 ns  |   1.7x  |

count_lt:

| n       | hit=5%  | hit=50% | hit=95% |
|---------|---------|---------|---------|
| 10      |  0.19x  |  1.24x  |  1.64x  |
| 100     |  0.53x  |  4.30x  |  6.14x  |
| 1000    |  2.62x  | 28.4x   | 49.9x   |
| 10000   | 19.7x   | 188x    | 356x    |
| 100000  | 157x    | 1512x   | 3101x   |

count_gt:

| n       | hit=5%  | hit=50% | hit=95% |
|---------|---------|---------|---------|
| 10      |  0.22x  |  1.19x  |  1.50x  |
| 100     |  0.62x  |  4.28x  |  6.22x  |
| 1000    |  2.70x  | 28.2x   | 48.9x   |
| 10000   | 20.4x   | 184x    | 353x    |
| 100000  | 157x    | 1488x   | 3077x   |

[meta-codesync]

@alibeklfc has imported this pull request. If you are a Meta employee, you can view this in D108627773.