fixed multipolygon error, submask speedup 1700x

[diogo-aos]

This pull requests makes 2 contributions:

• addressing the multipolygon problem
• speedup sub mask creation 1700x

This pull request addresses the multipolygon error in issues 9 and 12, by adding each individual polygon in the multipolygon.

The current sub mask creation is very slow because it uses Python loops. This function was reimplemented using NumPy native boolean logic. This implementation receives the category colors instead of the image size. Using cProfile, in our dataset, the speedup was of 1700x.

Contributions by myself and @fsmatilde.

[joeyst]

The changes both sped up generating annotations (while maintaining correct annotations) and fixed the multipolygon error for me as diogo-aos said. (P.S., thank you diogo-aos!)