Quantizer improvements

The original versions of these quantizers landed a couple months back,
and they were fine. Since then, we've had several opportunities to
iterate on them. This change lands improvements from those iterations.
TL;DR: 18% faster, big bug fix to WSMeans, code is cleaner (hopefully)

- QuantizerMap indexes an image's pixels, 'unique-ing' them by reducing
the pixel array to a map with keys of colors, and values of population.
This allows other quantizers to operate much more quickly: instead of
working on each pixel individually, they're able to operate in bulk.
- QuantizerWu uses flat arrays instead of 3D arrays and is more
understandable IMHO.
- QuantizerWsmeans has speed improvements, most importantly, it has a
big bug fix. When a Kmeans-based quantizer algo starts, it must first
assign the pixels to any one of the starting clusters. The original
implementation decided what cluster to assign a pixel to by finding the
cluster closest to the pixel. However, the algo terminates if no pixels
moved after one iteration of the algorithm, and since the pixels were
already in the cluster closest to them, the algo would immediately
terminate before it actually figured out where the cluster moved to
after pixels were assigned to it, and had a chance to move pixels around
based on that.
- Funnily enough, even though this _should_ mean Wsmeans got a lot
slower since it has to do more iterations, it is actually 16% faster

Additionally, during review of this CL:
An accidental dependency on iteration order of a Set was introduced,
causing inconsistent initialization of the mPoints array, creating
inconsistent results from the quantizer.

Removing the dependency on hashes of float[], and avoiding Maps
altogether, removes a dependency on hash codes of pointers that existed
during review, making it easier to have verifiable consistency across
iterations. This also improves speed slightly, from 55 ms to 39 ms
(tested on sunfish, first 9 wallpapers in Landscapes, City Scapes, and
Art categories, and averaged)

Bug: 189931209
Test: ran performance tests with VariationalKMeansQuantizer,
the previous Celebi = Wu + Wsmeans quantizer, and the new Celebi =
new Wu + new Wsmeans quantizers, over 100 iterations. Wu speed is
roughly the same, Wsmeans is 18%. Verified quantizer output is stable
for the same input pixels, run 100,000 times for each of two wallpapers.
Change-Id: I3324d29860c098ea1fd602b8d4197837e732f4f1