This came up while debugging memory corruption in the ngraph-tf tests: plaidml_ops_replicate was holding onto
a reference to a temporary that was destroyed while the reference was still live.  Replacing that code with a
call to get_argument() eliminated the temporary and fixed the issue.  And in general, there's no reason to be
allocating temporary vectors when we're pulling out a single argument, so this change replaces the other
sequences of get_arguments().at(), too.

The PrefixReshapeElimination pass doesn't work correctly if the network's operation's input and output shape
compatibility is being validated; an implicit broadcast operation is required to get the shapes right.  This
change moves the implementation back to PlaidML (since it's seeming less generally useful, and PlaidML has an
implicit broadcast operation), and fixes both it and the implicit broadcast pass to only apply when the
downstream operation is an elementwise operation (since PlaidML doesn't support automatic broadcasting for
contractions).

Leona/doc v0.23

Remove reshape file from op/util.

Add initial igpu docs and other important fixes for 0.22

fix spelling error

memory allocator for CPU 

Failing test on INTERPRETER backend.

Add QuantizedConvolution op to ngraph core

Removed deprecated fusion type constants.

Improve PlaidML backend option debugging

Migrate from r0.23

[MLIR] Bump MLIR repo to commit 82d5084 Wed Jun 26.