This particular implementation isn't 400x faster, nor 100x, nor even 10x, nor even 2x. It's about 20% faster, as small benchmarks and overall experimentation shows. E.g., looping a million times with old Decimal and new Decimal (in Crystal):

decimal old increment  13.20  ( 75.75ms) (± 5.93%)  61.0MB/op  54.68× slower
decimal new increment 721.82  (  1.39ms) (± 1.06%)    0.0B/op        fastest

But this is only a silly little benchmark. A simple loop 100000 times: [ ] in Novika now runs in about 17s instead of 20s, on my machine (Ryzen 3 2200g). The main thing is that we have 0 memory usage at most times now for numbers, while before even in simplest scenarios we had a small, but usage nevertheless.

This has almost no effect on the big picture, though. So for now, let this be is on hold, because it introduces complexity and possibility of bugs (the snake fix commit may be a bug in snake, or in the decimal implementation; the fact is that it has to do with random numbers, and there is no sure way to trigger it other than play snake for eternity).

Closes #1