Optimize amp for multi-gpu to enable FP16 gradients transfer across gpus

Fix float16 optimizer

init new amp, optimize inserting cast op for batchnorm

* remove unused inference_transpiler unit-tests

test=develop

* remove InferenceTranspiler usage in quantize_transpiler.py

test=develop

* Add LAMB optimizer

* Expose LAMB Optimizer's APIs

test=develop, test=document_preview

* Cleanup code & doc

test=develop, test=document_preview

* Update lamb optimizer's formula

test=develop

* init auto loss scaling

test=develop

* change API.spec

* change ifelse to switch and use reduce_sum to optimize checking isfinite

test=develop

* Remove redundant code

test=develop

* Init mixed precision training interface

* Add fp16 test script

test=develop

* All initializers support float16

test=develop

* Code cleanup & add more code annotations

test=develop

* Update API spec

test=develop

* Add usage example in doc

test=develop

test=develop

This reverts commit f5d5d7b2.

* Disable in_place in batch_norm API.

Fix str and bytes problem
Fix sorted problem
Fix math problem
Fix CI problem

* add test float16 inference accuracy example

* complete the test

* clean code

* add argument parse and refine tests

* add shell script

* add float16 benchmark code

* refine code

* prepare for contrib/float16

* put things in contrib float16 folder

* update benchmark result

* further update benchmark report

* add float16 inference report

* update report

* add float16 transpiler

* fix feed fetch target names mismatch

* fix cast op input change issue

* fix program desc flush error

* fix inconsistent var names in block desc bug

* code clean up

* add float16 infernce C++ example and fix prune bug

* remove share lod; it has already been performed in infershape

* slightly release test bound of test_image_classification

Fix a bug in save_inference_model and prune when the program is initailized by load_inference_model (#10011)

* Fix bug in save_inference_model and prune when the program is initialized by load_inference_program.

* Save the transpiled program instead.

In this way we can build and test using WITH_FLUID_ONLY flag being set
to ON.

- move paddle.v2.dataset,reader to paddle.dataset,reader
- remove unused code (which depends on v2) in paddle.dataset,reader

* fix is_test issue

* add paddle enforce

* fix bug

* add new func

* small fix

* address comments

* merge fluid dist tests

* update cmake

* Generated functions can take `*args` as inputs.

Use a new scope for inference in python unittest to avoid changing the value of variables for training.