LinearNetwork support for compensating for input/output synapse

[arvoelke]

There are ways to change the readout to account for extra filtering. This involves either assuming D != 0 or \dot{u} = 0 and then computing the derivatives of y needed to cancel out the additional filter (this can be done analytically from the state-space representation). For example, if the extra filter is Lowpass(tau), then the output would need to be (in the continuous case) tau*\dot{y} + y, where y = Cx + Du and \dot{y} = C\dot{x} + D\dot{u}.

[arvoelke]

Also note that the special case of C = I and D = 0 might be useful as a way of re-encoding state (using different encoders) or simply probing the state without any filter, but this becomes SIMO/MIMO.

[arvoelke]

Full support connected to #106.

[arvoelke]

The network's state.input field actually helps a fair way to this end (at least in the special case where the extra filter is equal to the recurrent synapse), since this gives x as a PSC. This can then be transformed with C in a biologically plausible way to give y without filtering.

[arvoelke]

Also note that it's possible to avoid needing du if u has already been filtered in which case we might not need the extra filter on this part.

[arvoelke]

Special case (for rolling window) of getting the state encoded into another ensemble while accounting for the extra filtering:

nengo.Connection(dn.input, X, transform=dn.B, synapse=dn.input_synapse)
nengo.Connection(dn.state, X, transform=dn.A, synapse=dn.synapse)

May also need to provide the same solver used by _make_core from dn.state.

For LinearNetwork can simply use dn.state.input?