[Feature Proposal] Add NaN and Inf checking to RL models

[hill-a]

As described in #358, It would be relatively simple to add invalid value detection to Stable-Baselines.

Adding CheckNaNTensorFlow to the losses, and use tf.add_check_numerics_ops in the policies, allows invalid value detection and prevents check_numerics from wrapping over evey operation in the graph (tf.add_check_numerics_ops does this by default).

Code example

class CheckNaNTensorFlow:
    def __init__(self):
        self.check_nan_ops = []
        self.old_ops = []

    def __enter__(self):
        # wrap list so we have a copy of the current list, as list are mutable
        self.old_ops = list(tf.get_default_graph().get_operations())  

    def __exit__(self):
        for op in tf.get_default_graph().get_operations():
            if not op in self.old_ops:
                try:
                    self.check_nan_ops.append(tf.check_numerics(op, ""))
                except TypeError:  # things like reshape or conditional operations are skipped
                    pass

with CheckNaNTensorFlow() as tf_nan:

    ...   # define the TF graph you want to check here

sess.run([val_1, val_2] + tf_nan.check_nan_ops, td_map)  # add to run

Additional context
This will probably add some overhead, my guess only as much as an if instruction. On CPU branch prediction should not bother too much of the performance, it might however hurt GPU performance significanlty, this needs verifing.

[hill-a]

PPO2, CPU, forward and back propagation without invalid value checking:

• 0.036 +- 0.01 ms

with invalid value checking:

• 0.036 +- 0.01 ms