fixed memory leak from opt return

[williamFalcon]

Fixes #1510
and a few other things along the way.

Now uses .detach to track metrics instead of .item() which keeps speed fast but decreases memory.

Also speeds things up:
Before this PR lightning slowed an epoch by 0.89 seconds, now only by 0.65 seconds (<1 sec per epoch). This baseline is compared with the same model trained using only PyTorch

[pep8speaks]

Hello @williamFalcon! Thanks for updating this PR.

• In the file pytorch_lightning/callbacks/model_checkpoint.py:

Line 114:111: E501 line too long (113 > 110 characters)

Comment last updated at 2020-04-19 20:23:02 UTC

[codecov]

Codecov Report

Merging #1528 into master will increase coverage by 0%.
The diff coverage is 92%.

@@          Coverage Diff           @@
##           master   #1528   +/-   ##
======================================
  Coverage      91%     91%           
======================================
  Files          67      68    +1     
  Lines        3782    3790    +8     
======================================
+ Hits         3437    3445    +8     
  Misses        345     345