Armory format: Pre-check for cracks within parallel section

[solardiz]

This is a controversial change for a slow format like this, but OTOH I don't want my decision not to make this format salt-only (where we'd have to check for cracks in crypt_all) to have any obvious performance cost, whereas the number of comparisons per crypt_all can be significant when there are many threads and SIMD vectors are wide (e.g., 192*8 in my recent test means 1536 comparisons, which when done sequentially out of crypt_all can have a measurable cost).

[solardiz]

Weird, this segfaulted on ci/circleci: arm, whereas the previous revision passed tests there. I wonder if I introduced or exposed a bug here, or if it's something intermittent - e.g., maybe we simply don't always have the same amount of RAM available there?

[solardiz]

I wonder if I introduced or exposed a bug here

Indeed, I did. Thesalt argument to crypt_all is allowed to be NULL, which I didn't handle. I could also reproduce the crash on Solaris/sparc32. Now I fixed that. However, I'm puzzled why this didn't show up on x86(-64) - is our self-test/benchmark on x86(-64) different, not ever setting salt = NULL there?