Bug fix: RM transitions

[tdelgado00]

Description

This pull request allows two states of a reward machine to be connected with multiple transitions with different rewards (as long as the different transitions have different formulas). In particular, since all terminal states are mapped to -1, this allows an RM state to transition to a terminal state with two different rewards depending on the formula that was satisfied. This solves a bug present in some of the environments.

This should close issue #50.

Changes

• Changed the transition functions in the reward machine class so that they are maps of (state, formula) -> new state/reward instead of (state, state) -> formula/reward, and updated the corresponding references.
• Added a set of terminal states to the reward machine class. With this, we don't map all terminal states to terminal_u anymore. This was necessary to make HRM find the right set of options.
• Tested the office gridworld multi-task and single-task. In the single-task case, task 4 (get the coffee and email and go to the office) isn't a good example of the difference between HRM and CRM. I added entry points for the 4 tasks in the single-task version, so they can be run by selecting env=Office-single-Ti-v0 with i=1, 2, 3, 4. The difference between CRM and HRM can now be seen for example with T1.