* randint API: remove out, devive, stop_gradient; add name; test=develop

* test=develop

* test=develop

* test=develop

According to paddle 2.0 standard
1, change flip api attr name 'dim' to 'axis'.
2, support empty axis
3, change example code to imperative mode.

According to paddle 2.0 standard.
1, chang api to def meshgrid(*args, **kwargs) we hide name argument in **kwargs.
2, add relate ut.
3, change example code to imperative mode

* experimental refactoring, test=develop

* add TranslatedLayer & remove StaticModelRunner, test=develop

* revert tracedlayer change, test=develop

* fix test_mnist unittest error, test=develop

* add doc & examples, test=develop

* polish doc details, test=develop

* add imperative.jit module, test=develop

* change TranslatedLayer pos, test=develop

* adjust jit module import path, test=develop

* polish doc based review result

* add SaveLoadConfig.separate_params to save paraams separately

* add Layer.buffer support, test=develop

* polish doc details based review result, test=develop

* polish details baesd review comments, test=develop

* add empty str check for param, test=develop

* add unittests, test=develop

* polish details based review comment, test=develop

* remove blanks in comment, test=develop

* polish doc details, test=develop

* update imperative doc link, test=develop

* add api attr for load, test=develop

* add api to clear intemediate tensors in analysis predictor. test=develop

* add python api. test=develop

* make default_collate_fn visible. test=develop

* test_dist_fleet_ctr disable, test=develop

[while grad]Support pruning op in find_op_path about while sub-block when appending backward (#25330)

Prune OPs which are not related with loss in while sub-block when constructing backward OP path.

Add Similarity Net as unit test. During the unit test, we found three problems:

1. The run_program_op has memory optimization error when running dy2stat net multiple times.
2. The support for SelectedRows can cause problem in dy2stat.
3. The return grammar has problem.

This PR fixes the 1. problem but modify codes for the 2. 3. problems to make PR smaller. I will fix those two problems in the next PR(s)

* Add the imperative quantization aware training.
*  This is the python part of Imperative QAT. test=develop

* fix softmax_with_cross_entropy cuda kernel overflow bug, test=develop

* replace old macro & for condition, test=develop

* polish details, test=develop

* add explicit specialization

* add skiplayernorm vector load if available

* test=develop

* polish pe exception process logic, test=develop

* fix unittest, test=develop

* add unittests, test=develop

* fix optimizer.state_dict and LRScheduler.state_dict to save/load dygraph,test=develop

* fix optimizer.state_dict and LRScheduler.state_dict to save/load dygraph,test=develop

* Add a judgment that state_dict/set_dict is used incorrectly,test=develop

* fix some doc error,test=develop

* fix current_step_lr for _LearningRateEpochDecay,test=develop

* remove some unsed code to improve coverage,test=develop

* remove some unsed code to improve coverage,test=develop

Use vector instruction (LDG.128) to improve qkv transpose. It
provides 1.4X speedup at same GPU base frequency.
test=develop

remove useless property

add new API:LambdaDecay

* attempt to resolve tls problem, test=develop

* add glibc version check, test=develop

* fix regex, test=develop

* refine get_libc_ver, test=develop

* refine get_libc_ver, test=develop

test=develop

* fix optimizer parameter is a iterator; test=develop

* fix parameter list None bug; test=develop

* use is not None; test=develop

* change list to iterable; test=develop