Fix antialiased line issues

[ianthomas23]

This is a WIP to fix various issues with antialiased lines for the impending release.

The major change is support for both 1 and 2-stage aggregations. Where possible (e.g. ds.any and ds.max) a single-stage agg is used on an individual line segment basis. ds.min always uses a 2-stage agg on an individual line segment basis. ds.count and ds.sum now have 2 options controlled by self_intersect boolean kwarg. If False the usual 2-stage agg is used which is very accurate but slower and ignores line self-intersections. If True, the default, a 1-stage agg is used that considers the previous line segment so that pixels are not overwritten at line segment joins; this therefore deals with self-intersections and is faster but there will be rendering artifacts in some situations (some are fixable, others are unavoidable).

I am open to suggestions for a better name for the kwarg than self_intersect.

Antialiased self-intersecting path using ds.count now looks like this (from 15-Large_Data.ipynb):

Antialiased multiple timeseries using ds.count now looks like this (from 3_Timeseries.ipynb):

[ianthomas23]

The above multiple timeseries plot has artifacts to the left of y-extrema. This is now fixed:

Here is the equivalent line_width=0 plot:

[ianthomas23]

Performance of antialiased lines is slower than non-antialiased for 2 reasons:

1. Low-level pixel-setting code is more complicated.
2. Two-stage agg code has multiple agg-combining calls, each of which is a NaN-aware combination of 2 arrays of same shape as canvas plus a reset of one of those arrays to zero (or NaN). This does not occur for one-stage agg.

Timings using code based on last plot in Timeseries jupyter notebook that combines nlines lines of 10-points each onto a 1000x400 pixel canvas. Timed line_width=1 for both ds.count(self_intersect=True) (1-stage agg) and ds.count(self_intersect=False) (2-stage agg) compared to line_width=0.

nlines	Slowdown factor for 1-stage agg	Slowdown factor for 2-stage agg
100	1.3	2.5
1k	1.6	4.7
10k	3.3	7.9
100k	5.4	13.0
1M	5.7	13.7

[ianthomas23]

Example output for multiple timeseries (1000 here) where x-values are spaced apart, for line_width=1:

Compare with line_width=0:

This uses agg=ds.count() and how='eq_hist'.

[jbednar]

I don't know any clearer option name, so it looks ready to merge to me!