Reduce cost of DAGCircuit.input_map and output_map accesses

[jakelishman]

          This recreates the entire dict (which could be 100s of qubits + clbits) on every lookup.

I actually expected it to have more of an impact on transpiler performance, but it seems like we only have a couple of transpiler passes that touch them - the disjoint utils and the "remove final X" passes. I think those are sufficient non-critical we can ignore it for this first PR, but it's definitely a performance pitfall we should rectify for user passes.

Originally posted by @jakelishman in #12550 (comment)

At the moment, the Rust-space input_map and output_map properties recreate Python dictionaries on every access, meaning operations like [dag.input_map[qubit] for qubit in dag.qubits] have suddenly become quadratic in the number of qubits, rather than linear. This doesn't bite us too much in practice in the transpiler immediately, but it's absolutely a performance pitfall waiting to happen in the current structure.