* CHERRY_PICK 18941, 18860: TRT fp16 support.

test=release/1.5

* CHERRY_PICK 19213: Fix BUG: Mask RCNN inference diff When using AnalysisPredictor.
    1. fix affine channel fuse pass.
    2. fix condition block op.
    3. fix merge lod tensor op bug.
    4. fix memory optim cause by reset lod op.

    test=release/1.5

fix mask rcnn
add interface for setting optim_cache_dir(eg: when in trt int8 mode, and load model from memory, there should be a interface for setting the trt calibration table data dir)

test=release/1.5

* fix: when use the load model from memory mode, the RAM occupy is high

test=release/1.5

test=develop

* add set_not_owned function for graph

* add scope set. test=develop

* add scope_ptr enforce not null before setting.test=develop

1. infernce multi card occupy
2. facebox model inference occupy too much
test=develop

refine trt first run
add quant dequant fuse pass
omit simplify_anakin_priorbox_detection template
omit transpose_flatten_concat_fuse template
test=develop

* modify the interface of Pass::Allay
test=develop

* Polish code
test=develop

* Fix Travis CI
test=develop

* fix Pass::Apply interface
test=develop

* Fix Travis CI
test=develop

test=develop

test=develop

test=develop

size_t is an unsigned integer, with a conversion rank
 larger than int, therefore in the following expression
 the int value was promoted to size_t, making it a
 subtraction of unsigned values. The result of such
 a subtraction is also an unsigned value.

1. graph and program_desc alignment
2. trt stream

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

* Initialize the elementwise plugin.

* Implement the basic CUDA kernel of elementwise plugin.
test=develop

test=develop

the parameters will load from CPUPlace, that will keep copying data
between CPU and GPU places.

test=develop