inflate: Read more bits when decoding

[klauspost]

When a block is not marked as the final block it should be possible to read 10 further bits ahead without breaking the promise of not overreading.

Fixes #231

Also write blocks with only EOF as TYPE 01 (predefined tables) with only an EOB literal. This saves 3-4 bytes at no cost.

The smallest block seems to be a predefined block with a single EOB, which would be 10 bits + EOB from current block (current limit).

This should make it possible to fill more bits at the time when decoding.

λ benchcmp old.txt new.txt
benchmark                       old ns/op     new ns/op     delta
BenchmarkGunzipCopy-12          27830317      26617762      -4.36%
BenchmarkGunzipNoWriteTo-12     27878505      26705660      -4.21%

benchmark                       old MB/s     new MB/s     speedup
BenchmarkGunzipCopy-12          171.50       179.31       1.05x
BenchmarkGunzipNoWriteTo-12     171.20       178.72       1.04x

[klauspost]

Maybe @dsnet can spot some obvious flaw. After all he had the great idea to extend it to the size of EOB.

If not, this would be a reasonable thing to upstream.

[klauspost]

Tried a few variations on reading. Didn't manage to find anything faster.