Created by: chenwhql

When run(startup_program) is executed before optimizer.minimize, because the parameters added by optimizer are not initialized, which will cause run failed. The error message is like:

this error occur to RunTimeInferShape phase. the program doesn't check whether the tensor in LearningRate variable is initialized, and get the dim of variable used to compare directly, which is the reason that this error messgae is not easy to understand.

So this PR add extra error message in related ops to give users a hint, the related ops are in paddle/fluid/operators/optimizers. I just add error hint in the first persistable var checking. The result after changed is like:

related changed op and corresponding optimizer：

• sgd_op: SGDOptimizer
• momentum_op: MomentumOptimizer, DGCMomentumOptimizer, LarsMomentumOptimizer
• adagrad_op: AdagradOptimizer
• adam_op: AdamOptimizer, LambOptimizer
• adamax_op: AdamaxOptimizer
• decayed_adagrad_op: DecayedAdagradOptimizer
• adadelta_op: AdadeltaOptimizer
• ftrl_op: FtrlOptimizer

no need to change op:

• rmsprop_op: RMSPropOptimizer

I have confirmed that all optimizer can trigger expected error hint when the exe.run(startup_program) is executed before optimizer.minimize. The unittests code: test_optimizer_vars_init.py