Speed up encode_rle function by 250-500x

[csaroff]

Converting a segmentation mask(691 x 1024 where pixel-intensity = segmentation class) to series of run-length encoded masks(label-studio format) was taking forever(about 95 seconds with 4 segmentation classes).

After improving the performance of encode_rle, the same conversion takes 174ms

Here's the performance differences with timeit:

[robhyb19]

@makseq This performance improvement that @csaroff made helped us immensely. Was wondering if there was an ETA on merging this PR in with master?

[makseq]

@robhyb19 I've checked it now, works well. Merged.