Add in place concat optimization.

Merge branch 'ayzhuang/in-place-concat' of https://github.com/NervanaSystems/ngraph into ayzhuang/in-place-concat

* m_direct_execution is used but not defined when NGRAPH_DEX_ONLY=TRUE

* keep the ifdef and move m_direct_execution out of the ifdef

Merge branch 'ayzhuang/in-place-concat' of https://github.com/NervanaSystems/ngraph into ayzhuang/in-place-concat

* Update ArgReduce to handle i64 indices.

* Formatting.

* Add throw for output types other than int32/64.

* Add output type to hash.

* Add type to throw.

* Interpreter doesn't currently support 64bit output indices for argmin/max and so disabling this test [JIRA:NGRAPH-3183].

* Implement partial shape/type propagation for Convolution; fail for want of unit tests

* Implement unit tests for partial shapes/types for Convolution

* fix Klockwork warnings CPU part 1

* fix spelling error

* fix a typo

* Fix gemm `input_c` broadcasting.

* Comments.

* Add comment

* [ONNX] Non-linear ops

* Style check

* Temp bn update commit.

* Add CUDNNBatchNorm which adds two additional outputs to batchnorm, the batch mean and batch inv variance.
The batch mean is the same as the output mean if the cummulative average factor is 1.0. Add BatchNormCache pass which replaces all BatchNorm ops that are inputs to BatchNormBackprop
with CUDNNBatchNorm which outputs the saved batch statistics directly to the backprop step.

* Updated bn cache pass, removed extra tests, added test checking that provided stats are used in bprop instead of batch stats.
This test was disabled for interpreter as the reference kernel needs to be updated to use provided statistics.

* Formatting.

* Update to new batch norm API.

* CUDNNBatchNorm -> BatchNormTrainingWithStats

* new line

* Preprocess input variance into BN denominator for cudnn (#1885)

* Add explicit cuda kernel to calculate what cuDNN describes as the inverse
variance. In reality, the backward cudnn kernel for BN requires 1.0f / sqrt(variance + eps),
which is the batchnorm denominator for each channel (a numerically stable inverse stddev).

This introduces op annotations for batch norm backprop and updates the cudnn_emitter to support the insertion of this cuda kernel when required.

* Disable second test on INTERPRETER.

* [CI fix, c.f. @chris lishka] Roll back #906 and #1853.

* disable some quantize tests on gpu.

* skeleton backend

* Code owner from if conditioning

* add simple placement for interpreter and register pass in hybrid

* placement policy applied

* clone the function if needed

* split the function

* Compile subfunctions in corresponding backends

* hybrid backed works as is for abc test

* cleanup

* add placement policy for CPU

* cleanup a little

* add simple op cost method to backend

* enable CPU pass via flag

* address clang-format PR issue

* reslove build

* clean-up

* update manifest

* disable HYBRID as default build

* style

* addressing offline discussion

* more offline discussion

* onnx: add information about a domain to operators set
Signed-off-by: Artur Wojcik <artur.wojcik@intel.com>

* onnx: updates after review
Signed-off-by: Artur Wojcik <artur.wojcik@intel.com>

* onnx: update comments in the code
Signed-off-by: Artur Wojcik <artur.wojcik@intel.com>

* onnx: fix bug in node's description method
Signed-off-by: Artur Wojcik <artur.wojcik@intel.com>

* onnx: fix CentOS compilation
Signed-off-by: Artur Wojcik <artur.wojcik@intel.com>

* onnx: more after review changes
Signed-off-by: Artur Wojcik <artur.wojcik@intel.com>

* Support for parameterized testing combined with backend driven manifests.

* clang-format fix

* Improved nGraph variants of gtest fixture and parameterized testing macros

* Reverted mistaken change to src/ngraph/runtime/gpu/CMakeLists.txt

* added half_toward_zero; all previous tests passing

* all rounding modes added with unit tests

* fix cpu emitter

* round mode doc

* round out round modes

* doc typo

* using  names for round modes

* use ceil/floor for rounding functions instead of round/nearbyint

* clean up doc

* equidistant

* Collapse dimensions for arithmetic reduction ops to support faster kernels

* Propagate in-place constants and allow in-place reshapes for more cases

* style fix

* Additional checks for parameter and constant to help backends that dont propagate in-place parameter and constant inputs

* Allow non-destructive pass through onlyu if memory sharing is disabled

* Address PR feedback

* Bug fix for collapse dimensions in case of null reduction

* split bn into bn_inference bn_training

* fix warnings

* Add GPU support for the new BN ops (#1569)

* Add GPU support and change batchnorm_globalstats test to use BNInference.

* Changed test back to using BNTraining for global stats and updated cudnn backend to account for it.

* Fix issues in merge with master.

* Formatting.

* CPU fixes

* remove 5-arg training BN for now

* more fixes

* python batchnorm changes

* fix onnx_import

* fix a call BatchNormInference c-tor

* yet another fix to BatchNormInference c-tor

* AND yet another fix to batchnorm_inference c-tor

* ops.py

* address adam's feedback

* Remove unnecessary parameter/argument.

* remove batch_norm_training_relu_with_global_stats

* remove bn_relu (training)

* Update runCI.sh

Accommodate manual reproduction script.

* Update README.md

* Update runCI.sh

* fixes
* added rebuild parameter

* Fix rebuilding nGraph

* Added execution permission to runCI.sh