Add dithering to (32b?) radial gradients

This is a dithering implementation for 32b.

It creates LSB changes in:
filltypespersp_8888
drawbitmaprect_8888
drawbitmaprect_gpu
gradients_8888
clamped_gradients_8888
filltypespersp_4444
drawbitmaprect_565
shadertext_8888
clamped_gradients_4444
drawbitmaprect_4444

The new radial-gradient GMs look correctly dithered, but still display banding.

There are no compile guards in here to protect old clients, but those can be added before commit if this implementation looks good.

[reed1, 2012-01-30 21:17:47]

lgtm

[TomH, 2012-01-30 22:26:22]

This will require about 50 rebaselines in chromium.
Also, it does not produce sufficient improvement of the appearance of the new GM
(http://codereview.appspot.com/5598048/).

[TomH, 2012-01-31 20:14:30]

To keep notes on gradient quality in one place:

 - 8b math appears to be the cause of off-by-one errors in the color table
(Build32bitCache()); there are 12 for the new GM - 4 in the primary halftable, 8
in the dithered.

 - After controlling for those errors, the RGB colors are rendered
monotonically, but there are still many off-by-one errors in the alpha channel,
e.g. ... 85 85 84 85 84 84 83 84 83 83 82 83 82 82 81 81 ...

[TomH, 2012-02-01 22:02:10]

This should be ready for commit (and GM rebaselining), but is parked while Mike
addresses more fundamental bugs we found in gradients.

[TomH, 2012-02-02 15:57:45]

Committed in r3126; will be rebaselined once some tool support is finished.