Add tests for BN_sqr() corner cases.

Test BN_sqr() with a newly allocated BN, a BN explicitly set to zero and
small values that fit in a single BN_ULONG.

Correct benchmark result computation on 32 bit platforms.

Use utime in order to make benchmarks less noisy and more consistent.

Set BN_FLG_CONSTTIME on benchmark inputs.

The public APIs still change behaviour based on BN_FLG_CONSTTIME - set it
to avoid benchmark noise.

Add benchmarks for BN_div()

Correct setup for BN_mul/BN_sqr benchmarks.

BN_rand() takes a bit length, not the top bit.

Rename bn_mul.c regress to bn_mul_div.c.