Revisit benchmark suite for QuantumCircuit.assign_parameters

[jakelishman]

We need to take a little care when modifying benchmarks, because any change invalidates the whole trace. If we're doing this (and we probably should), we probably want to take a todo issue to write more complete benchmarks of assign_parameters that test more of the input formats / combinations.

Originally posted by @jakelishman in #10792 (comment)

At the moment our benchmarking for QuantumCircuit.assign_parameters is not really indicative of the different typical access patterns. We should be testing the time taken to assign single parameters out of big lists in a loop (using the dictionary input), to assign all the parameters using the dictionary input and to assign all the parameters using an array input. We care about both inplace and new-circuit returns.

For different cases, we care about big "flat" circuits (think many rzs), but also about very deep composite circuits (think algorithm-like things with recursive structures within single compound instructions).