Fix the checked align-up helper for network sizes

Firmware code rounds sizes up with (n + a - 1) & ~(a - 1) on 32-bit unsigned integers. An AI called it standard, safe alignment math, but it is only correct when a is a power of two and the sum fits in 32 bits. The sizes now come from network input, so the replacement checks both and refuses anything else.

An assistant wrote align_up_checked(n, a). It is in the editor. It has a bug. Find it and fix it.

Input and output:

• n: the size to round up.
• a: the alignment.
• Return the smallest multiple of a that is at least n, or -1 when the request is not allowed.

Rules the tests check:

• n must be from 0 to 4294967295, and a from 1 to 4294967295. Otherwise return -1.
• a must be a power of two: 1, 2, 4, 8 and so on. Otherwise return -1.
• If n + a - 1 is greater than 4294967295, the rounded size does not fit in 32 bits: return -1.
• Otherwise return n rounded up to the next multiple of a, or n itself when it already is one.