Work around index-limiting bug in np.einsum

[Strilanc]

• Just turn on optimize when 26 or more indices
• See BUG: np.einsum accepts some subscripts only when optimize=True numpy/numpy#10926

Fixes #683

Example that works now:

import cirq
c = cirq.Circuit.from_ops(cirq.H.on_each(cirq.LineQubit.range(13)))
c.to_unitary_matrix()

I like that you can nearly see the individual operations happening, in system monitor, by watching the memory usage:

[cduck]

LGTM. Do you have a test for these cases?

[Strilanc]

No tests unfortunately; it's too expensive. By the time you have 26 indices you're doing 2^26 work per gate.