* Integrate APU

* fix code style issue

Summary: Reformat fp16 gemv, add FP16_MATMUL_WEIGHT

Differential Revision: https://phabricator.d.xiaomi.net/D144134

1. Support int32 input data type.
2. Support GatherV2 op
3. Add transpose to ExpandDim op.

1. input_data_formats flag in yaml supports OIHW
2. MaceTensor support OIHW data format

1. Remove data_format flag for ambiguous
2. add has_data_format for distinguish models without data format.

support kaldi-onnx model convert
support validate output data from url

Bring gain of top-1 accuracy (69.54% -> 70.20%)

1. Fix random device chosen bug.
2. Add random device chosen strategy for running model.

1. Convert OpenCL binary files to code.
2. example support OpenCL byte stream with --cl_binary_to_code flag.

add onnx_optimizer tool and update onnx info in docs

1. Abstact android and arm linux to one format
2. Support cross compilation for ARM linux
3. Related issue #36

1. Support memory optimization at runtime
2. Support memory type transformation automatically at runtime.
3. Move gpu winograd convolution to Conv2D.
4. Support Data Type transformation automatically.
5. MaceTensor API support optional data type.
6. Add input_data_formats and output_data_formats at yaml
7. Related issue(#363)

1. Move all files in kernels to ops
2. unify the op_def_registry and op_registry.
3. Support op clip: do not link the ops' code when the op is not registered.