Tweak heuristics for fast predictor selection

[veluca93]

Have you tried subtracting 1 iff nbits[0] <= (1/2)?

Even better I think could be to double all the bit costs (as that changes nothing), but remap bitcost 1 into 1 instead of 2 for 0, and perhaps bitcost 2 into 3 (or 4), or something like that.

Rationale being that RLE only kicks in for many consecutive 0s, so you'd need the 0 probability to be quite high.

Originally posted by @veluca93 in #19 (comment)