Make rlevel calculation more efficient, and power of 2 check more robust.

[feature-engineer]

See
https://cocalc.com/8324bfc9-d3df-4304-a29c-a5ccf0283797/raw/power%20of%202.ipynb

for comparison with old code.

[feature-engineer]

In case the benchmark file is not available, here are its contents:

import numpy as np
%alias_magic t timeit
def pow2_1():
    for step in range(1, 10000):
        if np.abs(np.log2(step) - np.round(np.log2(step))) > 1e-6:
            msg = "Row and column strides must be powers of 2."
            continue
        yield int(np.round(np.log2(step)))

def pow2_2():
    for step in range(1, 10000):
        rlevel = step.bit_length() - 1
        if (step % (1 << rlevel)) != 0:
            msg = "Row and column strides must be powers of 2."
            #raise ValueError(msg)
            continue
        yield rlevel

Performance test

%%t
levels1 = [rlevel for rlevel in pow2_1()]

69.4 ms ± 4.39 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

%%t
levels2 = [rlevel for rlevel in pow2_2()]

1.32 ms ± 20.7 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

Test

levels1 = [rlevel for rlevel in pow2_1()]
levels2 = [rlevel for rlevel in pow2_2()]
assert len(levels1 == levels2)
assert all(level1 == level2 for level1, level2 in zip(levels1, levels2))