This optimisation once again gives us nothing on realistic values
due to roughly uniform distribution.

BigNumCalculateMontParams previously did not check BigNumMod for memory
allocation failure anyway, which was thus harder to diagnose due to
the Montgomery Inverse of N still returning correctly.

Currently BigNumMod does not allocate memory for A words with leading
zeroes. For us words are 8 bytes. When doing RSA computations with
proper keys 8 zero bytes are nearly impossible and 16 zero bytes are
completely impossible. Avoiding this optimisation does not change
library behaviour for nearly all inputs but provides simpler guarantees
about memory usage and simplifies testing.

This reverts commit 7c695007.

Fix build