Use the popcount64c bit-counting sequence in Dar_WordCountOnes

[TrevorHansen]

This reopens #533, which was closed by accident when the fork it came from was deleted. GitHub cannot reopen a pull request whose source repository is gone, so this is the same change from a new fork, rebased onto the current master. The description below is unchanged from #533.

The popcount (hamming weight) instruction has been common on x86 processors since about 2008. It returns the number of 1-bits in a register with 3-cycle latency and 1-cycle throughput.

ABC uses hand rolled SIMD within a register (SWAR) to achieve the same.

static inline int Dar_WordCountOnes( unsigned uWord )
{
    uWord = (uWord & 0x55555555) + ((uWord>>1) & 0x55555555);
    uWord = (uWord & 0x33333333) + ((uWord>>2) & 0x33333333);
    uWord = (uWord & 0x0F0F0F0F) + ((uWord>>4) & 0x0F0F0F0F);
    uWord = (uWord & 0x00FF00FF) + ((uWord>>8) & 0x00FF00FF);
    return  (uWord & 0x0000FFFF) + (uWord>>16);
}

When passed the -mpopcnt flag, clang and GCC will pattern match and replace some matching SWAR with a popcount instruction.

Unfortunately, the SWAR above is not recognised by any GCC I tested (7 to 16), or any clang (9 to 22), so it emits 23 instructions instead of 1.

Popcounts are in other places in ABC (e.g. #436), so other users would benefit if the other uses were sped up. Only this function is in our hot-path - some CNF generation problems have 7% of their runtime in Dar_WordCountOnes, so I've focused on it.

The above function is the naive implementation (popcount64a) from:
https://en.wikipedia.org/wiki/Hamming_weight

This PR changes to the best implementation (popcount64c) from the same page:

uWord = uWord - ((uWord >> 1) & 0x55555555);
uWord = (uWord & 0x33333333) + ((uWord >> 2) & 0x33333333);
uWord = (uWord + (uWord >> 4)) & 0x0F0F0F0F;
return (int)((uWord * 0x01010101) >> 24);

Which is converted by GCC 10+ and clang 10+ to a single instruction, and is 8 fewer instructions if not compiled with -mpopcnt.