Drop code under #if HAVE_POSIX_MEMALIGN.

[zackw]

alg-yescrypt-platform.c contains code to use posix_memalign if mmap is
unavailable, but it’s not properly wired up to autoconf, so it never
actually gets used. I don’t think it’s worth keeping, because any
system that bothers to supply posix_memalign probably also has mmap.

I'd like @solardiz to comment on this change.

(This addresses one of the issues described in #30.)

[solardiz]

I support this change and am likely to make the same change upstream.

However, I disagree with part of the reasoning that "any system that bothers to supply posix_memalign probably also has mmap" - I guess Windows is a notable exception here, it just might not be relevant to libxcrypt. My own reasoning is that we maintain alternative source code (and struct fields to support it) anyway and it isn't slower - it's just some trivial extra calculations. We can accept this trivial runtime complication on systems that have posix_memalign in order to simplify the source code and improve the coverage of its testing.

[zackw]

I think I understand. Would you describe this as a better rationale for the change?

It is not worth attempting to use posix_memalign when mmap is not available, because the existing fallback to malloc + manual alignment is good enough for that situation.

(N.B. Current versions of Windows offer huge page allocations via VirtualAlloc, that might be worth looking into.)

[solardiz]

Yes, that wording for the rationale sounds better to me. And thanks for the info on Windows VirtualAlloc.

[zackw]

Rebased and pushed with the edited rationale.