[Feature request] Allow different network architectures for off-policy actor/critic

[araffin]

Currently, for on-policy algorithm we can specify net_arch=[dict(pi=[64], vf=[64])]

The idea would be to allow the same (expect the shared part which adds too much complexity) for off-policy algorithms:
net_arch=[dict(pi=[64], qf=[64])].

This should be fairly simple to implement.