lib: fix nearest_power() overflow handling on 32-bit platforms

[uwezkhan]

On 32-bit platforms, nearest_power() may return an undersized value
for inputs larger than the maximum representable power of two.

This can cause callers to receive a buffer smaller than requested when
rounding large allocation sizes.

Reject overflow-range inputs explicitly so nearest_power() never
returns a value smaller than the requested size.

This change preserves the existing nearest_power(0) == 1 behavior and
adds regression tests covering overflow-range inputs.