add concat and scale to paddle_use_bridges.h

* fix some error when compiling with mlu-sdk1.2.5

add cuda c++ doc

- 增加cuda c++ demo.
- 考虑到检测模型尾部一般是multiclass_nms，该kernel为host，如果fetch kernel为cuda的话，则会在此处插入无用的io_copy(host->cuda)，由于该原因，注释掉fetch的cuda kernel. 默认使用host的fetch kernel. 此处暗中进行的行为：每次predictor run完，都会默认把数据从cuda拷贝到cpu

* * support download bm_sdk, test=develop

* [LITE][BM] add slice op

* [LITE][BM] fix concat issue

* [LITE][BM] support reduce full ops, test=develop

* [LITE][BM] change test_resnet50 to change test_classify

* [LITE][BM] add cast op

* [LITE][BM] add reduce and other ops, test=develop

* [LITE][BM] add reduce,cast and other ops, test=develop

add cuda ci.

Organize cuda compilation products.

* add sgd op, test=develop

* test=develop

* add backend

* add kernel

* add grad kernel

* [NPU] subgraph's precision register to kAny

* [LITE][OPENCL] Change fp32 fc to fp16's. test=develop

* fix act in conv3x3opt opencl kernel. test=develop

* [LITE][OPENCL] fix opencl fc kernel. test=develop

* [LITE][OPENCL] add lws for opencl conv image kernel. test=develop