Probe.size_in is sometimes incorrect

[drasmuss]

Describe the bug

The Probe.size_in is always set to be the size_out of the object being probed. This is often correct, but the assumption fails for more unusual probes.

For example, when probing connection weights this will just report the size of the pre object, not the size of the weights. Or any probe on a learning rule will report a size_in of 0.

To reproduce

A minimal code snippet to reproduce the behavior:

import nengo
import numpy as np

with nengo.Network() as net:
    a = nengo.Ensemble(1, 2)
    b = nengo.Node(size_in=10)
    c = nengo.Connection(a, b, transform=np.ones((10, 2)))
    print(nengo.Probe(c, "weights").size_in)  # prints 10

    d = nengo.Connection(
        a, b, learning_rule_type=nengo.PES(), transform=np.ones((10, 2))
    )
    print(nengo.Probe(d.learning_rule).size_in)  # prints 0

Expected behavior

Probe.size_in should report the size of the data that will be recorded by that probe (e.g., (10, 2) for the weights probe, or maybe just 20 if we want to stick to integers).