ndarray benchmarks

The element-wise math ufuncs are benchmarked against NumPy's vectorized equivalents (same fixed-seed array to both, op run many times, results cross-checked) by scripts/numpy_compare.py. Each function's Performance row on the ndarray reference carries its own number. The table below is generated by that harness — see docs/performance.md for the methodology (striated rounds, medians, paired ratios) and the reference machine. The band column is the range of the per-round ratios, which on the large-array rows is wide enough to matter: a bare headline would hide that sqrt at 16384 lands anywhere from 0.6× to 1.1× depending on the round.