Previously, MR#approvals_before_merge is nil when it is <= project
(from `clamp_approvals_before_merge`).
This changed in f480d145,
which would always persist a min value.
This caused regression that updating MR would fail if project's
value is increased later on.

The relationship between value being nil and validation can be easily
overlooked.

Removing validation and taking the max of the two makes logic clearer.