1. Add a new attribute named 'startuo_seed' to RandomCropOp. If the input
'Seed' is not initialized, the 'startup_seed' will be used to replace
it.

2. Refine CustomReader. Add a member variable 'scope_' to it. The
'scope_' will act as the global scope of preprocesing, making it
possiable to save something cross batches.

* "add some api reference"

* "fix based on preview"

* "fix evaluator"

* "remove template doc"

* "add doc for exec"

* "add more changes"

* "fix based on preview"

* "chagne code format"

Give assign layer's second parameter 'output' a default value: None. If
it is None, the output variable will be created inside the layer.

* Make the CUDA kernel of concat correct and fix unit tests.

* Add sub_blocks of lr_decay_op to pserver_prog after distribute_transpiler

* Remove unused logs and logics

* 1. Add ops to new block (considering the nested block condition)
2. Follow the original hierarchy of blocks
3. Change the function's name and remove debug lines