[WIP] Use user-defined NumPy dtypes for AutoDiffXd, Expression, etc.

[EricCousineau-TRI]

Resolves #8116

See tracking issue for progress.

\cc @soonho-tri @RussTedrake

---

This change is 

[EricCousineau-TRI]

@RussTedrake The main backwards-incompatible change this introduces is that np.array(..., dtype=Variable) can no longer use .dot(...), since the NumPy C API does not seem to permit returning different types for dot products (unfortunately):
617ce19#diff-81cc7ecb26961a59c2c5ecf90bcfe8faR31

Is that OK?

[RussTedrake]

yuck, but I understand . if you think that this is the best outcome, i can live with it.

[EricCousineau-TRI]

I do believe the other benefits outweigh this cost, but I've submitted numpy/numpy#10876 as a request.

[EricCousineau-TRI]

@jwnimmer-tri This PR has an initial incantation of building NumPy from an upstream PR and using the resulting wheel in a build.
I've also submitted RussTedrake/underactuated#113, which just needed the dot product adjustments.

Do you think I should lump the NumPy repository stuff all together in this PR, or make a separate one?
(My preference is the same one, but I can separate it out if it's less to deal with.)

EDIT: Er, drake_cloc had reported what looked like ~600 LOC of changed, but this definitely indicates more. I'll split it out, then.

[EricCousineau-TRI]

+@jwnimmer-tri for feature review, if you have time.

---

Review status: 0 of 45 files reviewed at latest revision, all discussions resolved.

---

Comments from Reviewable

[jwnimmer-tri]

Even setting aside the four commits that are from #8608, this PR seems quite large. Perhaps it makes sense just do to "autodiff: Use user-defined dtype" (+ the pybind11 bump) as the next PR in this train? (Or symbolic instead of autodiff, if that was better somehow.)

[EricCousineau-TRI]

Yuppers, can do!

[sherm1]

@EricCousineau-TRI -- last activity on this PR was in May. Do you want to keep it open?

[EricCousineau-TRI]

Yes, I would like to keep this open. Once deadlines settle down here, I'd like to finish this up, especially in light of MultibodyTree, and supporting autodiff (and eventually symbolic) for them.

[SeanCurtis-TRI]

@EricCousineau-TRI assigned to you for PR hygiene.

[EricCousineau-TRI]

Closing for now. Can re-open later.