MRG, ENH: Speed up convert_forward_solution

[larsoner]

I noticed in https://travis-ci.org/mne-tools/mne-python/jobs/650978029?utm_medium=notification&utm_source=github_status that a lot of time was spent in:

115.44s call     mne/inverse_sparse/tests/test_mxne_inverse.py::test_mxne_inverse_standard

Profiling suggests that about half the time was spent in convert_forward_solution, and a lot of that was in _normal_orth, so this might be a speed regression due to #7290.

In any case, this PR takes that test from 23 sec to 14 sec on my system. Timings for mne/tests/test_chpi.py don't change, so the n-dim generalization that this makes use of seems to be not so painful for the single vector case.

[codecov]

Codecov Report

Merging #7333 into master will decrease coverage by 1.87%.
The diff coverage is 100%.

@@            Coverage Diff            @@
##           master   #7333      +/-   ##
=========================================
- Coverage   89.88%     88%   -1.88%     
=========================================
  Files         450     450              
  Lines       81264   81264              
  Branches    12918   12918              
=========================================
- Hits        73046   71520    -1526     
- Misses       5396    6981    +1585     
+ Partials     2822    2763      -59

[larsoner]

I think ultimately we are suffering from numpy/numpy#15580, but this should still be merged since it's useful in any case

[larsoner]

(and the failures are fixed by #7325, even if the timeouts are not)