Improve docs for big data

[astrofrog]

As mentioned in #15 by @keflavich regarding big_data.rst:

"If we're going to identify a method that is more efficient, we should illustrate how to use the more efficient method. I think this calls for a refactor of the _apply_numpy_function method to use @ChrisBeaumont's slab-by-slab method if axis is passed and add a memmap=False kwarg in the _apply_numpy_function & inherited tools. I guess this ought to go to a separate issue."

[ChrisBeaumont]

I can implement this. However, I don't think I understand what memmap=False is meant to communicate. Does this specify whether the computation happens slabwise? If so, I think we should use the same convention as moment (how='cube'|'slice'|'ray')

[ChrisBeaumont]

Along these lines, we might want to add a quick tutorial discussing what a scientist needs to keep in mind to work productively with large cubes. I see lots of advice scattered through the docs (which is good), but we might want to have the equivalent of "getting started with big cubes" section

[ChrisBeaumont]

78 Adds the suggested tutorial, and partially implements slicewise-reduction for numpy methods.

There is a bit more that could be done in this regard:

• Implement ray-wise reduction
• Extend slicewise reduction to argmin and argmax, which need a different algorithm

[astrofrog]

Most of what is done so far is fine for 0.1, re-scheduling this for 0.2 for the remaining issues.