Cluster train doc update

[putcn]

Since the work mode between trainer and master changed from "master actively find trainer and deliver training job" to "master passively listen to trainer's request for task and deliver training job", this PR is for the design doc change according to the implementation update.

[helinwang]

Thank you! Very appreciated!!!

[helinwang]

CI unit test failed:

 [13:16:19]	.E [13:16:19]	====================================================================== [13:16:19]	ERROR: test_uniform_random_gpu (__main__.UniformRandomTest) [13:16:19]	---------------------------------------------------------------------- [13:16:19]	Traceback (most recent call last): [13:16:19]	File "test_uniform_random_op.py", line 13, in test_uniform_random_gpu [13:16:19]	self.uniform_random_test(place=core.GPUPlace(0)) [13:16:19]	File "test_uniform_random_op.py", line 28, in uniform_random_test [13:16:19]	ctx = core.DeviceContext.create(place) [13:16:19]	RuntimeError: basic_string::_M_construct null not valid [13:16:19] [13:16:19]	---------------------------------------------------------------------- [13:16:19]	Ran 2 tests in 0.510s 

Can you maybe try rebase / pull to the latest develop branch code (possibly contains the fix to the problem).