* Add moving average absmax op in quantilize-aware training.

* Support Sync Batch Norm.
* Note, do not enable it in one device.

Usage:

build_strategy = fluid.BuildStrategy()
build_strategy.sync_batch_norm = True
binary = fluid.compiler.CompiledProgram(tp).with_data_parallel(
        loss_name=loss_mean.name,
        build_strategy=build_strategy)

* Impl fp16 compute kernel for slice_op

test=develop

* Use data() to replace mutable_data()

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

2. remove lod check in reshape op
test=develop

* Add cpu_quantize_squash_pass for C-API quantization

test=develop

* add cpu_quantize_squash_pass teste

* fix lint: add include memory unorderd_map and unordered_set

test=develop

* lint fix 2

* fixes

test=develop

* refactored

test=develop

* fix windows ci

test=develop

test=develop

* resolve #15618
Backgroud: the PR #15398 raised the box_coder op performance regression, we optimized the code via the more efficency leveraging opemmp.

test=develop

* fix broadcast with mp mode

* polish code test=develop

* fix bcast strategy test=develop

* fic cpplint test=develop

* fix py3 failed test=develop

* fix comment test=develop

* update comment test=develop

* remove concast_cast and refactor code test=develop

* reduce flag use test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

* Remove Desc in Forward Pass

* Refactor VarBase

* Add dbg info

* Only check type in imperative mode

* Polish code and support optimizer

test=develop

* Fix stop gradient problem in PyLayer

test=develop