ENH: implement lazy imports for optional dependencies

[neutrinoceros]

This is another step to resolve #27
Overall, this gets us to a state where core dependencies (numpy + sympy) are responsible for at least 50-60 % of unyt's import time.

This refactor is similar to my work in yt yt-project/yt#3935
It allows for optional dependencies to never be imported before they are actually requested by users, reducing unyt's own startup time when heavy optional dependencies are installed.

Making matplotlib a true "on-demand" import took a little more work, but now import unyt doesn't cause matplotlib to be imported too when available, reducing the startup time by about 15% to 25% with respect to unyt 2.9.2 (on my machine).

[neutrinoceros]

this now includes changes for dask, recently added with #185 (ping @chrishavlin)

[neutrinoceros]

this is now based on #272

[neutrinoceros]

@jzuhone @chrishavlin can I request your reviews on this ?

[jzuhone]

Curious about the doctests bit, but otherwise looks good.

[chrishavlin]

Looks good! Left a couple minor comments.

One additional question -- why'd you put the unyt_arrayConverter class in _mpl_array_converter.__init__.py instead of a new module, _mpl_array_converter.py ?

[neutrinoceros]

One additional question -- why'd you put the unyt_arrayConverter class in _mpl_array_converter.init.py instead of a new module, _mpl_array_converter.py ?

I can't remember. I would guess I went through several iterations and that form was the first one with which I managed to pass tests, for some reason. Now I agree that a module is better than a mono-module package, and I've checked that it does not in fact break anything, so let's roll with that.

update: it seems that the simpler form breaks on Python 3.8 (see https://github.com/yt-project/unyt/pull/250/checks), so I'll revert this change