clarification of environments used in A2C algorithm

[TomorrowIsAnOtherDay]

Currently we are using envname&NoFrameSkip-v4 for training.

PARL/examples/A2C/a2c_config.py

Line 21 in ee3e8dc

'env_name': 'PongNoFrameskip-v4',

We should add more details in READEME for users to make sure that they use the actual environments.

[TomorrowIsAnOtherDay]

note:

PARL/examples/A2C/atari_model.py

Line 17 in d18f19a

from paddle.fluid.param_attr import ParamAttr

this line should be removed.

[TomorrowIsAnOtherDay]

note2:

Currently the command tensorboard cannot be found after installing PARL in Mac, and we should add fix this issue in our installation script.

[TomorrowIsAnOtherDay]

note3:
refine the API get_weights of parl.Model so it return a dict like parl.Algorithm and parl.Agent.

[TomorrowIsAnOtherDay]

note4:
use a CPU version of Paddle for checking the compatibility on CPU

[zenghsh3]

note5:
when creating too many actors: "zmq.error.ZMQError: Too many open files"

[TomorrowIsAnOtherDay]

All the environments use gym API now. I