fix the gradient bug for the topk v2 

* add WeightedRandomSampler. test=develop

* Add checkpoint to quantize

* Change bfloat16 option

* add mkldnn inplace op version test

* update mkldnn_inplace fuse pass

* update the inplace test

* fix test_multinomial

* fix test_multinomial add 0 prob

* DataLoader support not auto collate batch. test=develop

PaddleSeg uses tuple as parameter of transpose in dygraph code:
https://github.com/PaddlePaddle/PaddleSeg/blob/release/v0.7.0/dygraph/paddleseg/models/danet.py#L152

However, in dy2stat, static code doesn't support the perm as a tuple. This PR fixed it.

* Fix gradients with ignore_idx in softmax_with_cross_entropy.
test=develop

* Fix gradients with ignore_idx in softmax_with_cross_entropy on cpu.
Remove softmax_with_cross_entropy from op_threshold_white_list.
test=develop

* Fix test_softmax_cross_entropy_op.py.
test=develop

Fix no_grad setting in AdamW.
test=develop

* Support squeezed label as input in paddle.metric.Accuracy
* Revert cifar and fix UT

* register op_version for addto

* upgrade pass capability

* change eq to le

* change eq to le

* fix merge

* fix vgg num classes

* fix summary for rnn gru lstm

* open a part of GPU unittest for windows

* open a part of GPU unittest for windows

I found the unittest failed due to batch size. Maybe the reason is that our CI machine has limited memory. I decreased the batch size.

* update, test=develop

register the op version for some ops

this is a preemptive measure, the unit test was failing occasionally with
cudnn internal error

* add onnx export module, test=develop

* add unit test for paddle.onnx.export

* adjust api & doc

* fix some typo

* make Numpy version is below 1.19.3

* fix 2.0 doc

also fix `no_grad` alias

* skip_layernorm_op done

* add unittest

* slice op convertor support trt < 6

* skip_layernorm only work in ernie

As the title

fix the add_n english doc