new feature for profiling training runs (#782)

* initial implementation

* formatting, pass through profiler, docstring

* call profiler during training

* add initial tests

* report stats when training is done

* fix formatting

* error handling, bugfix in passthroughprofiler

* finish documenting profiler arg in Trainer

* relax required precision for profiling tests

* option to dump cProfiler results to text file

* use logging, format with black

* include profiler in docs

* improved logging and better docs

* appease the linter

* better summaries, wrapper for iterables

* fix typo

* allow profiler=True creation

* more documentation

* add tests for advanced profiler

* Update trainer.py

* make profilers accessible in pl.utilities

* reorg profiler files

* change import for profiler tests

Co-authored-by: William Falcon <waf2107@columbia.edu>

docs/source/common-cases.rst

@@ -13,9 +13,15 @@ gradient clipping
 modifying training via hooks
 =============================
 
-
-
 .. toctree::
    :maxdepth: 3
 
-   pl_examples
+   pl_examples
+
+
+profiling a training run
+========================
+.. toctree::
+    :maxdepth: 1
+    
+    profiler

docs/source/profiler.rst

@@ -0,0 +1,10 @@
+.. role:: hidden
+    :class: hidden-section
+
+
+Profiling performance during training
+===========
+.. automodule:: pytorch_lightning.profiler
+   :exclude-members:
+        _abc_impl,
+        summarize,