add collective op for cpu using gloo and paddle.distributed.* apis

add the top v2 for the paddlepaddle api 2.0

test=develop

* add cudnn related strategies to DistributedStrategy

* add unique_v2 op

* remove unique_v2 op

* update doc

test=develop

* Fix the cmake-function named inference_download_and_uncompress with Windows, test=develop

* Fix some problems when remove limit of unittests on Windows, test=develop

* Using URL to download file instead of DOWNLOAD_COMMAND. test=develop

* optimized transformation form tensor to numpy, test=develop

* optimized transformation form tensor to numpy, pass pre-commit, test=develop

* modify fetchophandle zerocopy to deepcopy in PE&CUP, test=develop

* modify py:array construct, test=develop

* fix _fetch_var to use deep copy, test=develop

* Enhance affine grid operator:
1. Add cuda kernel
2. Add align corners options
test=develop

* Move new affine_grid api to functional
test=develop

* Add CUDA kernel for affine_grid.
test=develop

* Add more unitest for grid sample API
test=develop

test=develop

* api2.0 paddle.nn.Bilinear and paddle.nn.functional.bilinear, test=develop

* api2.0 fix code examples, test=develop

* modify test_bilinear_api, about place,to_tensor , test=develop

* re pass pre-commit, test=develop

* Update common.py

* fix BilinearTensorProduct ci error, test=develop

add op_function_generator.exe retry in windows

This PR enhance CPU kernel and add new CUDA kernel to make grid_sample support:

- align_corners: with bool type.
- padding mode: which can be in ['zeros', 'reflect', 'border']
- Interpolation mode: which ca be in ['bilinear', 'nearest']

The old CPU and CUDNN version only support align_corners=true, padding_mode='zeros' and interpolation_mode='bilinear'.

The behavior of the new version op in default mode is compatible with the old version.

change to use bce_loss op, add numel check for bce_loss.

* add div, floor_div, remainder

add the isnan, isfinite, isinf api for the paddle 2.0

* add the new api and op for the argmax, argmin