Matrix gain results in cryptic error with encoders

[hunse]

Describe the bug

Creating an ensemble with a gain/bias distribution that results in a matrix (e.g. nengo.dists.Choice([[1.0]])) results in a cryptic build error (because the encoders end up being a 3-tensor):

  File "workspace/nengo/nengo/builder/operator.py", line 539, in reshape_dot
    "shape mismatch in %s: %s x %s -> %s" % (tag, A.shape, X.shape, Y.shape)
nengo.exceptions.BuildError: shape mismatch in <Ensemble (unlabeled) at 0x7fdb36501f98> encoding: (100, 100, 1) x (1,) -> (100,)

To reproduce

import nengo

with nengo.Network() as net:
    u = nengo.Node([0])
    a = nengo.Ensemble(
        100, 1, gain=nengo.dists.Choice([[1.0]]), bias=nengo.dists.Choice([[0.0]])
    )
    nengo.Connection(u, a)

with nengo.Simulator(net) as sim:
    sim.step()

Expected behavior

Ideally, we should catch this before build. Either way, the resulting error should clearly point to the problem.

We could also consider fixing this so the case where a 1-D vector is passed to Choice (as above) works.

[drasmuss]

This has come up a few different places recently, we should add some better validation for Distributions in general (e.g. sample something from the dist and check the shape in the same way we do if a direct ndarray is used).