Use 16x unrolling for generic floating point squared euclidean distance computations

[vekterli]

@baldersheim please review. Noticed this when looking at the generated ARM NEON machine code.

This lets the auto-vectorized code use all 4 32-bit lanes of a 128-bit SIMD register instead of just 2, which doubles performance compared to double (which is in line with what could be expected).

[vekterli]

After some experiments, a 16x unrolling factor seems to be optimal for both float and double. Further increases either has no effect or reduces performance.

Before:

double : sum=2610046000000.000000 of N=1000000 and vector length=1000 took 589
float  : sum=2610046000000.000000 of N=1000000 and vector length=1000 took 570
int8_t : sum=2610046000000.000000 of N=1000000 and vector length=1000 took 82

After:

double : sum=2610046000000.000000 of N=1000000 and vector length=1000 took 121
float  : sum=2610046000000.000000 of N=1000000 and vector length=1000 took 81
int8_t : sum=2610046000000.000000 of N=1000000 and vector length=1000 took 82