Add support for dict generator input_fn in learn_io (#7045)

* Branch to test_tensorflow

* using types.GeneratorType for validation

* changed to receving generator function

This allows to run multiple epochs of a finite generator

* using placeholders as keys to feed dictionary

* added correct counter for the index_placeholder

added corrections to tests

* removing extra spaces

* doc_string fixes

* Changed logic to handle final batch

if the generator ends  it restarts to fill the final batch

* Added check that dictionary is returned from gen

fixes in tests

* Fixes in docstrings, tests and target_key bug

* All tests fixed and pass

* Added generator_input_fn to __init__

* revert repo and branch in dockerfile.devel to original

* docstring changes based on reviewer comments

* target returns now a dict of labels

target_key can now be a list of keys to be used to populate the target

* added more test functions for the new label dict possible errors

* if label is a single feature return a tensor (to make it work with estimator object)

moved asserts to begining of generator_input_fn
added test for dictionary with single input as generator return

* Added sorting to keys to ensure order

* moved _GeneratorFeedFn to contrib

* fixed numpy import

* Revert "Added sorting to keys to ensure order"

This reverts commit 4ad3a0ddee7f263c650fefe9139a95099d457467.

# Conflicts:
#	tensorflow/python/estimator/inputs/queues/feeding_functions.py

* removed duplicate generator feed function from contrib feeding_functions.py

* indendation fixes and added sorted keys again

* fixed imports

* removed extra lines from feeding_functions

* removed unnecessary import in feeding_functions

* fixed sorting of keys

added ability to pass any container of target_keys (e.g. list set tuple)
raise KeyError if mismatch between genarator dict keys
docstring fixes

* add empty line between class methods

* Fixing some final formatting

* Undo misguided types import

* fixed indentation back to two spaces