Alpha filter

[hunse]

Add Alpha filter implementation

[drasmuss]

Made various fixups, more details in the commit messages. Let me know if those changes look good to you, then will merge!

[hunse]

The problem with just using one step in test_lowpass_trainable is that all the system really needs to learn is the initial filter state, it doesn't need to learn tau. I set trainable=False for tau and found that this test still passes.

I would recommend using multiple steps and checking explicitly that tau has changed just to make sure it's training properly, as I did in the modified test. Unfortunately, that probably means using looser tolerances, because once it's optimizing over multiple timesteps it will choose a larger initial_value so that after the initial value decays for a few timesteps, it's still near the target value of 1.

It would be nice if there was a way we could have really tight tolerances and be sure that tau is training properly. The only way I can think to do that would be to have the target be a lowpass-filtered version of the input, with e.g. a tau that's twice as long as the initial tau. That way, the target should be perfectly learnable. (In the alpha filter test, I just use all ones, which it can never learn perfectly since it would need an infinite tau; hence the looser tolerances.)

[drasmuss]

Updated the test in the other PR (as it worked better after those changes), see 9decbd0

[hunse]

This should be ``cell.get_initial_state`` or something like that.

[hunse]

The new test looks good. So I'm fine to merge this as-is (with the minor fix to the docstring I mentioned above), and then the test will get fixed when we merge #5.

Also, it might be a good idea to double-check the docstrings for the base classes, since I obviously didn't do a great job updating them when I reduced the code redundancy.