avoid usize overflow when sizing scratch buffers in yescrypt_body

[dxbjavid]

The scratch buffers in yescrypt_body are sized with 32 * (r as usize) * (n as usize) and friends. r/n/p come from a parsed $y$ hash, and on 32-bit targets that multiply wraps even though the u64 validation just above passes: classic-mode n=65536, r=32768 makes 32*r*n reach 2^36, wrapping to 0 so V becomes vec![0; 0] and smix1 writes past it (debug builds panic on the overflow). Compute the lengths in u64, already bounded by the checks, and convert with usize::try_from like the existing n cast.