Add partial draft of run-time parameters RFC

[jakelishman]

Apologies, there's quite a lot more I intended to write (and refine the current text) about further parts of the design and my progress, but I now need this to be up and available for others immediately.

I have pushed the branch move-instruction-parameters to my fork which contains what I have currently done. Roughly, this is:

• added most of the backwards compatibility shim for Instruction.params
• added a parameters field to CircuitInstruction, DAGOpNode and DAGDepNode (the last two might not be complete - I don't think I had fully finished updating DAGCircuit and DAGDependency)
• updated the CircuitInstruction context class to insert a reference to themselves into contained Instruction instances
• added a first pass at defining a "parameter specification" for Instruction instances
• added a first (poor) pass at a new Instruction.decompose interface to replace Instruction.definition. What I've done isn't right at all - my failure here is what caused me to push this from Terra 0.21. The starts of the design in the RFC about defining "input parameters" for QuantumCircuit are the right way forwards here, and the weird attempts currently in the branch to handle Instruction.decompose and Instruction.definition should be completely redone.
• made the initialisers of some gates accept zero parameters - in the new preferred form, the rz instruction will be initialised as RZGate(), and it's only when we add it to a circuit that we define its operands.

Certainly still to do:

• the "parameter specification" probably needs tweaking (I would keep everything about this private for Terra 0.22, because we'll want to refine the system heavily for typed dynamic parameters)
• all uses of Instruction.params (whether through circuits or DAGs) in Terra need updating to the new form.
• QuantumCircuit needs to learn how to specify input parameters, so we can define the new interface for Instruction.decompose (or Instruction.definition).
• The new Instruction.decompose / Instruction.definition method needs to be in place so ad-hoc no-subclass Instruction instances can be created and parameterised correctly (e.g. the return value from Instruction.inverse, etc).

Please don't take any of this RFC / any of the code I wrote in my branch on Terra as gospel - a lot of the interfaces doubtless need tweaking or rewriting.