Plaidml

This release includes:
* Preview EDSL / plaid2 support
* Initial MLIR integration
* Several workarounds for AMD & iGPU Metal compiler issues
* Fixes for several shape related issues
* New keras backend functions contributed by the community

Other notes:
Stripe based backends should exceed v0 backends for most targets. Enable experimental devices and set USE_STRIPE=1 to test.

See Changelog in README.md

* New Metal HAL
* New CUDA HAL

Release Notes:
* Now supports ONNX 1.1.0 as a backend through onnx-plaidml
* Preliminary support for LLVM. Currently only supports CPUs, and only on Linux and macOS. More soon.
* Support for LSTMs & RNNs with static loop sizes, such as examples/imdb_lstm.py (from Keras)
* Training networks with embeddings is especially slow (96)
* RNNs are only staticly sized if the input's sequence length is explicitly specified (97)
* Fixes bug related to embeddings (92)
* Adds a shared generic op library in python to make creating frontends easier
* plaidml-keras now uses this library
* Uses plaidml/toolchain for builds
* Building for ARM easy (–-config=linux_arm_32v7)
* Various fixes for bugs (89)

* Adds new op library and revamps python interface

Fixes several Keras issues
Adds preliminary Windows support