argon2: detect allocation failures in hash_password_into

[newpavlov]

Closes #566

[jonasmalacofilho]

Note that my allocation-only benchmark in #566 suggests this could be a tiny bit slower than vec![Block::default, ...].

So maybe it would be good to benchmark this with high-ish memory sizes, or at least with the 19 MiB and 46 MiB sizes recommended by OWASP?

[tarcieri]

Maybe we should try to land #547 first so we can gauge any impact on performance?

[newpavlov]

Maybe we should try to land #547 first so we can gauge any impact on performance?

I think we should merge this PR, then #547 without the alignment workaround after it was migrated to Blocks, and then finally add the alignment workaround to Blocks in a separate PR. I am still not 100% convinced that we should add the workaround. It needs additional investigation in my opinion.