Round instead of truncating pixel number when using Agg

When rasterizing a figure the allowed figure sizes are effectively

discretized by the finite dpi. Previously, we simply truncated when

computing the number of pixels to use, now we use `numpy.round` which

will avoid floating point issues when setting the size of the figure

in inches to get a specific size of in pixels.

Additionally, the Agg based backends will adjust the size of the

figure to match the actual size (given by the number of pixels and the

dpi) during the draw process.