Branch 134608953 (#4645)

* Make input_feature_key optional for users who pass in a custom signature_fn that does not use the first arg (examples).
Change: 134479303

* Add a fused batch normalization op, supporting both cuDNN and an Eigen-based
CPU implementation.
Change: 134479333

* Update ops-related pbtxt files.
Change: 134482005

* Update generated Python Op docs.
Change: 134482366

* Move contrib/quantization ops to tensorflow/core
Change: 134501895

* Update ops-related pbtxt files.
Change: 134503035

* Update generated Python Op docs.
Change: 134503299

* Automated rollback of change 134501895
Change: 134506649

* Add seek to FileIO via RandomAccessInputStream and BufferedInputStream,
by introducing a Seek() API to those implementations.  Seek is different
than SkipNBytes in that you can Seek backwards, whereas on a stream,
you can only move forwards.  While Seek() on RandomAccessInputStream
sort of makes sense, Seek() on BufferedInputStream makes a little less
sense.  I avoided adding Seek() to the InputStreamInterface because I didn't
want all implementors of InputStreamInterface to have to implement Seek(),
only those for which it made some sense.  However, I did have to add
a Reset() API to InputStreamInterface to allow reseting + skipping as the
method of seeking backwards.

Adds read(n) to FileIO.  Defaults to 'read to the end of the file' as
usual, and allows reading n bytes.

This change is to help enable FileIO to be used as a fileobj for gzipfile,
among other things.
Change: 134542845

* TF Checkpoint V2: make CheckpointReader work with the V2 format.

If the same checkpoint prefix identifies both a V1 checkpoint and a V2
checkpoint on disk, the V2 version takes priority -- which matches the same
behavior as the RestoreV2 op.

Typical usage:

$ bazel run tensorflow/python/tools:inspect_checkpoint -- --file_name=<V2 ckpt prefix>

Other changes: add DebugString() and Contains() to BundleReader.
Change: 134543092

* Don't show dropdown options if there is only 1 possible option.
Change: 134547683

* Adds a method to resample tensors from of batches at either a per-item specific rate, or at a rate proportional to per-item weights.
Change: 134548493

* Backwards-compatible updates to TF Distributions interface; mostly
surfacing subclass-private-specialization docstrings.

* subclass specializations now have signatures in Distributions class
  (self-describing interface)
* subclass docstrings of private methods are now properly appended
  to the public methods (i.e., Normal.log_pdf.doc += Normal._log_pdf.doc)
* modified places that return AttributeError to return the right type of error
  (either NotImplemented, TypeError, or ValueError)
Change: 134550708

* Add sequence_length static shape checks to tf.nn.(dynamic_)?rnn.
Change: 134551966

* Automated rollback of change 134503035
Change: 134560875

* go: Add an example.

Add an example (that will appear in Go doc) for a real use of the Go TensorFlow
APIs - using a pre-defined image recognition model for inference.

While at it a couple of minor tweaks:
- NewSession now accepts 'nil' for options as the documentation says it does
- Convenience accessors for the Outputs of an Operation
- Ability to extract (possibly partial) shapes from an Output

Another step towards #10
Change: 134560938

* Add license header to file.
Change: 134564823

* Simplify and speed up shape inference code in tf.contrib.distributions.Categorical.
Change: 134566685

* Update generated Python Op docs.
Change: 134567587

* tfdbg: Debug session wrapper framework

The design is done with supporting the following use cases in mind:
  1) A CLI debugging a locally running, in-process session.
  2) A CLI debugging a remotely running session.
  3) TensorBoard visual debugger debugging a remotely running session.

The wrapper provides the abilities to
  1) Intercept a run() call to a wrapped session and insert debug tensor watches according to debug URLs.
  2) Release control to an external object before and after the run() call, so that the external object can perform actions such as launching a UI to allow users to inspect the intermediate tensors and partition graphs from the run() call.
  3) (To be implemented in a future CL) Intercept a run() call and give control to DebugStepper to perform stepping / continuing-to actions on the graph.
  4) (To be implemented in a future CL) Enter an instruction loop to let an external object (e.g., remote client) launch run() and cont() calls remotely.
Change: 134575545

* Include new RestoreV2 op in android_extended_ops_group2.
Change: 134576245

* Move mutex when closing Android TF session to prevent deadlock.
Change: 134576424

* Add log loss metric for use in TensorForest, which is used for some datasets like Criteo.
Change: 134578604

* Centralize selection in Projector class. Add SelectionContext, an interface
that registers and dispatches selection events, to break a circular dependency
between Projector and ScatterPlot. Remove selection listener from ScatterPlot. Remove 'points()' private property on vz-projector.ts. Add neighbors to selection awareness.
Change: 134583197

* Automated rollback of change 123445810
Change: 134595813

* Fix typo in estimator docs.
Change: 134602527

* Update TensorBoard TAG to 30.
Change: 134603748

* Add color legend.
Change: 134604465

* Update generated Python Op docs.
Change: 134604749

* Append the timestamp to each /individualImage URL, so the refresh
button causes the browser to reload the image when it changes.
Change: 134605115

* Add tests for distribution_util.
Change: 134608953