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.

* Convert some ops to use Output<Node> inputs

* Remove duplicate validation

* Some GetOutputElement changes to help with Output<Node>

* Review comments

* Deallocate all temp tensors before return

* style-apply

* remove unique_ptr from allocator static allocation

* - for default allocator return address of the static object on stack instead of returning ptr to dynamically allocated object

* Revert "- for default allocator return address of the static object on stack instead of returning ptr to dynamically allocated object"

This reverts commit 126b269cb298e6eb8b99b1b9a90d5aa7cf8fb948.

* use raw pointer in set_host_memory_allocator

* fix build error

* Finish moving things into test/backend

* Update CODEOWNERS for moved file

* Tweak GPU manifest for renamed test

* [MLIR] Fix naming convention in MLIR files

Add naming convention note per file to state which files should use
nGraph naming convention and which MLIR naming convention and align
naming convention in those files with such a note.

* Remove m-prefix

[CMAKE] Change CODEGEN error on windows to be more grammatically correct

Leona/doc v0.24

[MLIR] Enable Gather Op

Add nbench support for double buffering

Remove support for adjoints of nodes

Python capsules support for ngraph::Function

* Pad: add pad mode SYMMETRIC

* Style.

* Disable UT on unsupported backends.

* Make required numpy version explicit.

* Read setup requirements from requirements.txt

Restore DEFAULT_FLOAT_TOLERANCE_BITS logic in lrn, sigmoid, softmax. Should allow re-enabling PlaidML tests (#3277)

* Added rpath setting for PlaidML backend

* Fix colons

* Separate build and install rpath

* Explicitly set target properties