Refactor RB code to use the new Clifford operator class and update the CNOTDihedral class

[ShellyGarion]

Summary

Refactor RB code to use the new Clifford operator class in Terra quantum_info and the updated CNOTDihedral class:
close #391
close #392

This code allows:

• More efficient RB.
• Do not need pre-generated tables (pkl files) for the 2-qubit groups (Clifford / CNOT-Dihedral)
• RB on the Clifford group for more than 2 qubits.

Details and comments

• Add a new class: RBgroup that handles the group methods needed for RB
• Remove the files: Clifford.py, clifford_utils.py, dihedral_utils.py, basic_utils.py and test_clifford.py
• Refactor circuit.py so that it will use RBgroup class
• Refactor rb_test.py to include tests for RB on the Clifford group with more than 2 qubits

API updates needed in randomized_benchmarking_seq:

• instead of:
  interleaved_gates: a list of lists of gates that will be interleaved (Optional[List[List[str]]]),
  the input will be:
  interleaved_elem: a list of group elements (Cliffords / CNOTDihedral objects).
• add rand_seed: random number generator seed

[ShellyGarion]

The overall structure looks good, but there are a lot of inefficiencies in the Dihedral group method using nested python for loops, if these were rewritten to use Numpy array and array methods (dot etc), it should give performance improvements.

I agree with your comment. Indeed, the current CNOTDihedral class has many for loops, and does not use numpy arrays.
However, it this PR I haven't changed the code of the existing CNOTDihedral class, and only added some extra functionality.
I suggest to merge this PR (since there are other PRs that depend on it), and then I will open another PR for rewriting the code to improve the performance of the CNOTDihedral class.

[ShellyGarion]

@chriseclectic - thank you very much for the detailed review. I've fixed your latest comments.
Do you think that the code can be merged now?
(the test failure is due to build timeout and not related to this PR).

[chriseclectic]

Looks good, we can address the Numpy performance issues in a follow up PR