Reset op resets to initial signal value

[tbekolay]

Motivation and context:
When I was preparing the 3.1 release, I found that a slow test was failing. To fix it, I added a Reset op, which required resetting the value to a two-dimensional value instead of a scalar. Rather than deal with that, I though it made more sense to have Reset reset signals to their initial values, whatever they may be.

In the end, I went with another approach to fix the slow test failure, but this PR still seems like a useful one to have IMO.

How has this been tested?
Existing tests cover it.

How long should this take to review?

• Quick (less than 40 lines changed or changes are straightforward)

Types of changes:

• New feature (non-breaking change which adds functionality)

Checklist:

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

Still to do:

• Discuss if this change is worthwhile
• Add changelog entry

[hunse]

One downside to this for a backend like NengoOCL is it means we'll have to either store all the arrays of initial values on the GPU, or copy them over each timestep. Neither seems ideal.

I'd probably suggest a hybrid approach where for Reset ops that are resetting the whole signal to the same value, we specify that value (e.g. Reset(sig, value=0)), but we also allow for a way to reset to initial value in the case where we need that (e.g. Reset(sig, value=None) or something like that).