Fix memory explosion when auto-calculating canvas range extents with dask

[jonmmease]

This PR improves the Canvas autorange logic to avoid bringing each full x/y array into memory in order to compute the min/max values.

This fixes #668 for me. Here is the test case I I've been using to diagnose and test this PR. I started a Dask distributed instance on a large workstation and persisted the ~3-billion point OSM dataset into memory. After this persist, the Dask dashboard reports ~28GB memory used. Then I set up a VM with 8GB of RAM and connected it as a client of the distributed scheduler. I then performed a cvs.points aggregation without specifying x_range/y_range, causing the autorange logic to be invoked.

import dask.dataframe as dd
import datashader as ds
osm = dd.read_parquet('/path/to/osm-3billion.parq/').persist()
cvs = ds.Canvas()
agg = cvs.points(osm, x='x', y='y')

Before these changes, the memory usage of the client would climb steadily until the kernel died. With these changes, the aggregation completes successfully with no noticeable increase in memory usage on the client.

cc @jacobtomlinson

[jbednar]

Looks great, thanks!