Fixed bugs and inconsistencies in benchmarks

[s8phsaue]

• fixed several occurences in method 'unsafe' where 'self.state' instead of the argument 'state' was used
• adjusted 'max_episode_steps' check to greater equal (for consistency)
• removed 'self.unsafe' call in 'step'. The reason here is that the original code calls 'env.unsafe' directly after therefore making it redundant. Furthermore, by removing the call to 'unsafe' we can explore algorithms that use, e.g., a cost limit greater equal than 1.
• fixed road2d's and noisyroad2d's 'unsafe' method to use velocity instead of position