It would be better to make process_function as argument of run instead of the constructor

In some situation, we'd like to train model in multiple phase with different process function! Current interface design we have to create different trainer, and install the handlers at each trainer. I suggest to make the process_function as an argument of run instead of the constructor, that we can train as below:

trainer = Engine()
# adding checkpoint handlers, logging handlers

# the process function in different training phase
def proc1(engine, batch):
    pass

def proc2(engine, batch):
   pass

# training with different process function sequentially and sharing the engine

trainer.run(proc1, data1, max_epochs=10)
trainer.run(proc2, data2, max_epochs=10) # and we can also use different data

Start by reading the Engine class and its current constructor and run interfaces, then trace how process_function and handlers are stored across runs. The change is done when sequential runs can use different process functions and datasets while sharing one Engine and its handlers, with tests covering that workflow.