Add differentiation benchmarks

[brocksam]

This PR:

• Creates a new benchmark module for benchmarking SymPy's differentiation functions.
• Introduces two new simple differentiation benchmarks based on the "baby example" and the "lighthouse example" from Griewank & Walther (2008).
• Introduces a new large differentiation benchmark based on the Gait2D model from Ackermann and van den Bogert (2010) and Gerritsen et al. (1998), and in this GitHub repo by Moore . See also this related GitHub issue.

This PR introduces so test_ methods in addition to the time_ methods that will be run by asv. These are included here for bookkeeping with the hope that in the future SymPy's benchmarking will support a wider range of benchmark types (see this SymPy mailing list thread and this asv GitHub issue).

[moorepants]

One thing about asv is that the benchmarks shouldn't take much time. Does that 2d walking model take long to compute? How to you plan to handle benchmarks that take a long time if so? See #8.

[moorepants]

LGTM