Refactor synapse delay

[arvoelke]

Motivation and context:
Fixes #938 by correcting the behaviour in #1535 when D != 0. This is accomplished by refactoring the delay in the synapse, such that y = Cx + Du is set at the beginning of the time-step and x = Ax + Bu is updated at the end of the time-step.

In #1535 both equations were being updated at the end of the time-step, which caused them to be applied in the wrong order in the case of D != 0. This implements the correct set of equations for a discrete-time LTI system, and is consistent with old behaviour when D = 0.

Compared to the current master branch, systems that have D != 0 and len(A) > 0 no longer have a two time-step delay.

Another improvement with this PR is the Simulator produces the same results as applying filt, which in turn are both consistent with scipy.signal.lfilter. This was not the case before (regardless of D).

More details can be found here:

• Off-by-one in LinearFilter.General step #938 (comment)
• Off-by-one in LinearFilter.General step #938 (comment)

Interactions with other PRs:
Rebased onto #1387 (relies on having the state of the synapse being a signal), which is in turn rebased onto #1535 (state-space simulation of synapses).

How has this been tested?
Added test from #938 (comment) which now passes.
Rest of testing is a work-in-progress.

How long should this take to review?

• Lengthy (more than 150 lines changed or changes are complicated)

Where should a reviewer start?
Work-in-progress.

Types of changes:

• Breaking change (fix or feature that causes existing functionality to change)

Checklist:

• I have read the CONTRIBUTING.rst document.
• I have updated the documentation accordingly.
• I have included a changelog entry.
• I have added tests to cover my changes.
• I have run the test suite locally and all tests passed.

Still to do:

• Investigate failing tests
• Profile speed differences
• Investigate how this interacts with graph optimizer
• Investigate whether a custom operator is absolutely necessary
• Update usage of synapse=0 to synapse=nengo.Delay()
• Test Lowpass.filt difference with synapse = 0 versus synapse = nengo.Delay() and with / without y0 != 0 (Lowpass.filt raises shape error when tau == 0 and y0 != 0 #1590)
• Update documentation and changelog
• Add test from State space synapses #1535 (comment)
• More tests that are sensitive to the corrected behaviour (as discussed in State space synapses #1535 and Off-by-one in LinearFilter.General step #938)

Feedback would be helpful in determining what is worth doing.