They are aggregated using a strength expression

* replaces mat / scalar with mat * scalar**-1.
* replace list argument to subs with a dict

(Although there are pending changes to SymEngine for the above two
issues, they aren't released in a version yet.)

* Removed deep as a kwarg to expand(). It doesn't do what we want
anyway.

* Disabled caching for tests in test_kernels. This exercises SymEngine
code generation every time the tests run. The execution time for these
tests without caching is about 6 minutes for me, so the overhead is
acceptable.

Closes #16 on gitlab.

This changes the P2P stage to be aware of self interactions in the FMM
and to exclude them, if desired. Self interactions are specified by
passing an extra array called *source_to_target* which matches source
numbers to target numbers.

These expansions use information from the PDE to compress the
number of needed Taylor coefficients.