Created by: typhoonzero

Reason of this change:

1. IR passes need to modify startup program if the pass add/rename/delete variables, and the variable needs an initializer.
2. Passes are executed by ParallelExecutor, and we want to run statup after passes.Apply, so we need to run startup in ParallelExecutor constructor.
3. To keep API compatibility, we can allow user to call executor.run(startup) before ParallelExecutor constructor
4. User may need to load a model before running, need to be compatible of loading models

Cases

Consider below cases, this change should work on all of them:

# Normal cases
exe.run(startup)
pe = fluid.ParallelExecutor()
pe.run()

# single node load model and fine tune
exe.run(startup)
fluid.load_vars()  # or fluid.load_persistables etc.
pe = fluid.ParallelExecutor()
pe.run()

# single node load model and fine tune
exe.run(startup)
fluid.load_vars()  # or fluid.load_persistables etc.
pe = fluid.ParallelExecutor()
pe.run()

# single node simplified
pe = fluid.ParallelExecutor() # run startup inside here
pe.run()

# single node load model
fluid.load_vars()
pe = fluid.ParallelExecutor() # run startup inside here
pe.run()

# distributed mode
main, startup = fluid.DistributedTranspiler().transpile(...)
exe.run(startup)
fluid.load_vars()
pe = fluid.ParallelExecutor() # run startup inside here
pe.run()