API is synchronous per thread and threads are coordinated so that
we know when we hit the last iteration everything is done.
Using join() to gate end of iterations was introducing too much
overhead to timing as verified via checking traces.

* Re-enable -Wswitch and -Wswitch-enum

* Collapse identical switch cases

* Fix (I think) compilation errors

* Partially fix missing cases in GPU xformer, so clang will show me the rest of them...

* One last enum value

* Rename PlaidML_Executable::save -> save_as_format

* Repair regression in PlaidML tensor impl

This was caused by the recent removal of the offset parameter for tensor read/write operations -- we missed a
spot where read/write were being called for synchronization purposes.

* Disable a few more PlaidML tests pending triage

* Skip elision of reshape->reshape

It turns out this doesn't work, because the downstream reshape's input_order axis vector is incorrect if the
upstream reshape is removed.

* Add element type to PlaidML tensor debug output

* Use nGraph booleans for PlaidML boolean data

We'd previously been using i8; that's been deprecated for boolean data now that we have an explicit boolean
element type.

* Set PlaidML convolution output shapes correctly

We weren't transposing the output shape; we were computing the right data, but the incorrect shape metadata
causes validation to fail.

* Add a PlaidML implicit broadcast op

Better nGraph shape validation was tripping up PlaidML's use of a reshape to replace explicit broadcasts with
implicit NumPy-style broadcasts (since the reshape's output shape would be incorrect for the downstream
elementwise operation).  Adding this implicit broadcast operation lets PlaidML tell nGraph something useful
about the shapes, making validation pass (when it's otherwise correct).

* remove tensor offset from tensor read/write calls because it was never used

* fix build errors

* fix build errors

* fix python test errors

* more python fixes

* revert change

* Make old version of read/write deprecated

* fix python read overload

* one more try to fix python binding

* fix python

* yet another try

* why is this so hard

* fix?

* add text to changes.md

Avoid negative values in int64 initialization for cases where int64 parameters are used as indices (#3004)

* update visualize tree file extenstions and output formats

* fix runtime error

* remove throw from dtor (#2854)

* add special handle for the 'NNP_xx' op name

* style

* add special handle for the 'NNP_xx' op name

* style

* use description as suggest by bob

* Fix clang compiler warnings

* Remove unintended file.

* style

* Not part of PR

* Another extra closure ref

* More warnings from merges

* Lambda arg was used

* Refactor distributed to be isolated to a few files.

* Fix type-o

* style

* review comments

* type-o

* typo

* Return name

* Change warmup to run full, normal iterations

* style

* rework PerformanceCounter class

* Update intel gpu to new PerformanceCounter and enable performance on interpreter

* Add ConstantToBroadcast pass

* address review comment

* fix merge error

* constant/io sizes

* add sizes and counts for all data io in a json

* Add details to model sizing

* change filemode

* change filemode

* fix filemode

* Add NodeInput and NodeOutput classes

* Deprecate Node::get_inputs, Node::get_outputs, Node::get_output_inputs. Remove Node::get_input_from and Node::get_output_from

* Privatize most fields of Node

* Make deprecation of descriptor-munching classes optional

* Review comments

* Adapt ReshapeSinking to use raw pointers for NodeInput

* Fix ZDTE (thought I had already done in this branch, weird); style

* wip

* Change get_node_outputs() and get_node_inputs() to return vectors

* Updates after merge

* Whoops, forgot to define these functions

* {NodeInput,NodeOutput} -> {Input,Output}

* Kill shared_ptr in Output

* Move Input and Output into node.hpp

* Templatize the underlying node (sub)type in Input and Output

* Eliminate some get_input_* and get_output_* functions

* Change get_outputs and get_inputs back to their original names; rename NGRAPH_DEPRECATE_IO_DESCRIPTORS to NGRAPH_DEPRECATE_OLD_NODE_APIS

* Miscellaneous cleanup

* More cleanup

* Unbreak CPU build

* Simplify unit tests

* Make Node less friendly

* Deprecate more get_output_* and get_input_* functions

* A couple of PR comments

* Make the deprecation stuff more generally available

* Better comment

* Be more consistent about [] vs. at

* Set rpath for mkl-dnn.

* Set library rpath to  on Linux.

* Use patchelf to set rpath of prebuilt mklml library.

* Add patchelf to Linux Dockerfiles.

* Revert "Add patchelf to Linux Dockerfiles."

This reverts commit 1769505a866061552942e19467ddcc0dad0922e8.

* Revert "Use patchelf to set rpath of prebuilt mklml library."

This reverts commit 726f6553a0450520328607177d64baf48fa93dd2.

* Copy cldnn runtime.

* Copy mlsl libraries.

Update TBB build script for Windows.

Fix typo.

Fix incorrect omp lib name on Windows.

Fix incorrect tbb.dll path on Windows.

Make LIBRARY and ARCHIVE output directory consistent.

Function missing on Windows.

Update test::util::all_close() to fix compilation issue on Windows

Export CPU_Executable on Windows.

Change nbench path for unit-test on Windows.

Change copy to copy_if_different.

Install CPU backend on Windows.

Disable tools test on Windows.

Disable two failing unit test on Windows CPU.

Fix incorrect CPU backend install path on Windows.