Decomposition of fused ops in PlaidML backend

Remove CPURuntimeContext mkldnn primitives and workspaces assignments…

Add PlaidML Quantize & Dequantize

Enable ONNX Runtime CI check on nGraph PRs

Fix fused sigmoid overflow.

Update CONTRIB.md

Add DynReplaceSlice op; make Dimension support negatives

Add DynElimination for DynReshape

use lock_guard to protect the access to m_exec_map in cpu_backend

[Fused] RNN and  GRU cells