* 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

2. remove lod check in reshape op
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

* 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

* Revert "Revert "Add Event for TensorCopy" (#16022)"

This reverts commit e2da3a5b.

* use default stream
test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

Make parent_idx a dispensable output for beam_search op to support models saved by older paddle version. (#16106)

test=develop

test=develop

test=develop

1). add static trt load model
2). fix bug: when device_id is not 0, the trt will have a bug
test=develop

add simple static mode loading
test=develop

test=develop

1. bind trt input and output to fluid tensors

Make parent_idx a dispensable output for beam_search op to support models saved by older paddle version. (#16106)

test=develop

test=develop

- Support setting attr in config
- Support setting dtype and initializer for input in config
test=develop

* polish the cast op doc

test=develop

* follow comments

test=develop

* fix api.spec

test=develop