widen s_cost * p_cost to usize when sizing parallel memory

[dxbjavid]

With the parallel feature, hash_password_into and balloon_m size the per-thread memory as s_cost * p_cost evaluated in u32 and only then cast to usize. Both come from NonZeroU32 params and can be parsed from a hash string, so a product reaching 2^32 wraps: release builds get an undersized (often zero-length) buffer and a passing MemoryTooLittle check, debug builds panic on the multiply. Casting each operand to usize first computes the real size on 64-bit targets.