Merge changes from topic "nnapi-cherrypick-from-master-to-aosp"

* changes:
  Fix nullptr check
  Fix input shape in LSTM Eval for Sequence LSTM.
  Update API to allow collecting execution duration.
  Add a new ResultCode corresponding to ErrorStatus::DEVICE_UNAVAILABLE
  Added ANEURALNETWORKS_TENSOR_QUANT8_SYMM type.
  Add BIDIRECTIONAL_SEQUENCE_RNN op
  Add AHardwareBuffer support for NNAPI
  Implement NDK interface for NNAPI compilation caching.
  Clean and reorder operation enums.
  Add spec for DETECTION_POSTPROCESS.
  Fix a typo in the unidirectional LSTM doc
  Add quantized implementation of GENERATE_PROPOSALS.
  Add new op GENERATE_PROPOSALS.
  Mark ANNModel_setOperandSymmPerChannelQuantParams as introduced in Q
  Fix typos in the docs
  NNAPI Burst -- runtime V1_2 skeletal implementation
  Refactors LSTM Eval for reuse in Uni(Bi)directionalSequenceLSTM.
  NNAPI Burst -- NDK interface
  Fix DynamicOutputShapeTest.box_with_nms_limit_dynamic_output_shape*.
  Adds UNIDIRECTIONAL_SEQUENCE_RNN op
  Add quantized implementation for ROI_POOLING.
  Implement NDK interface for dynamic output shape.
  Treat CPU as a Device in StepExecutor
  TRANSPOSE_CONV_2D with per-channel quantization
  Adds missing OperationValidationTest tests.
  Refactors LSTM step function for reuse.
  GROUPED_CONV_2D with per-channel quantization.
  Add quantized implementation for BOX_WITH_NMS_LIMIT.
  Add new op BOX_WITH_NMS_LIMIT.
  Add quantized implementation for HEATMAP_MAX_KEYPOINT.
  Modify HEATMAP_MAX_KEY_POINT spec.
  Add quantized implementation of AXIS_ALIGN_BBOX_TRANSFORM.
  Add quantized implementation of ROI_ALIGN.
  Regenerate tests with the new test generator.
  Add new OperandType TENSOR_QUANT16_ASYMM.
  Add NDK API to allow developers to query the device type
  Refactor RNN op
  Regenerate test codes with the new test generator.
  Test dynamic output shape in GeneratedTests.
  Let CpuExecutor return correct ResultCode on fail.
  Implement dynamic output shape in CpuExecutor and Sample Driver.
  Remove kTypeNames and kErrorNames
  Implement IDevice::getType in SampleDriver
  DEPTHWISE_CONV_2D with dilation factors
  Split weights and bias parameter for quantized LSTM
  Remove scratch buffers and redundant output from quantized lstm
  Per-channel quantization support in DataTypeConverter.
  Update ROI_POOLING spec.
  Update ROI_ALIGN spec.
  Add OperationValidationTest for QUANTIZE
  Fix OperationValidationTest.CONV_2D_quant8_per_channel
  Support dynamic output shape in 1.2 HAL interface.
  Modify operation enum.
  Regenerate tests.
  Simplify AXIS_ALIGNED_BBOX_TRANSFORM op.
  Add float16 support for QUANTIZE
  Add float16 support to LSTM op
  Add float16 support for POW op
  CONV_2D with dilation factors.
  Remove kOperationNames