Avoid configuring pulse defaults in circuit assembly unless needed

[mtreinish]

Summary

This commit addresses a performance regression in assemble() introduced
in #6167. That PR added the ability for users to configure pulse
properties via assemble kwargs (such as lo frequencies). While this
should really only be exposed via a Backend's options object and terra
shouldn't have to deal with it, for legacy/backwards compat reasons we
need to expose these options via assemble(). However, that PR was a bit
aggressive in when it needed to parse the pulse defaults and properties
to determine if input was valid. The intent there was for circuits jobs
that users were configuring pulse configuration we'd be providing the
same checks as with pulse jobs. However for circuits jobs this
additional validation is only needed if the user has specified a pulse
configuration options and we need to check that they're valid. If a user
doesn't we shouldn't be touching anything pulse, especially as just
loading pulse defaults can result in a large performance regression
(roughly 2 orders of magnitude over assembly without it). To fix this is
commit adjusts the logic when we parse the pulse defaults (and all the
subsequent logic that uses them) to be if it's a pulse job, or if it's a
circuit job and one of the named pulse paramters has been set. This
returns the performance for pure circuits job to what it was prior to
PR #6167 but still retains the functionality that #6167 provided when
it's needed.

Details and comments

[mtreinish]

Running:

    import time

    import qiskit
    from qiskit.circuit.library import QuantumVolume
    from qiskit.test.mock import FakeLegacyMontreal

    qubits = 7
    shots = 8192
    memory = False
    pulse = False
    circuits = 1
    backend = FakeLegacyMontreal()
    # Pre-cache propertie
    backend.properties()
    qv_circuit = QuantumVolume(qubits, seed=42)
    qv_circuit.measure_all()
    circs = qv_circuit
    experiments = qiskit.transpile(circs, backend=backend, seed_transpiler=42)
    assemble_start = time.time()
    qobj = qiskit.assemble(experiments, backend=backend, shots=shots, memory=memory)
    print(time.time() - assemble_start)

The assemble runtime goes from: ~0.36427 seconds without this PR to ~0.00523 seconds with this PR

[jakelishman]

Looks good to me.