[WIP] Runtime improvements

[gemeinl]

Reduced runtime of TUHAbonormal from 2 min on master to 44s.
Reduced runtime of TUH from 58 min on master to 10 min.

master:

this PR:

[codecov]

Codecov Report

Merging #321 (12792fb) into master (9cc8c68) will increase coverage by 0.06%.
The diff coverage is 98.11%.

@@            Coverage Diff             @@
##           master     #321      +/-   ##
==========================================
+ Coverage   83.03%   83.10%   +0.06%     
==========================================
  Files          55       55              
  Lines        3838     3853      +15     
==========================================
+ Hits         3187     3202      +15     
  Misses        651      651              

[gemeinl]

Despite the improvements introduced with this PR, load_concat_dataset is still quite slow.
Thereby, close to 100% of the consumed time falls on mne.io.read_raw_fif, so I guess there is not much we can do.

A way faster solution for saving and loading our datasets would be to use pickle.
I am not sure whether we should include / recommend that though...

What do you guys think? Any ideas for further speed ups and improvements?

[agramfort]

can you also compare to torch.save ? I think it's what most pytorch users would expect. And the .pt extension is clear of how to load such files.