Compute curvature on mesh boundary and use in model specifications

[emmetfrancis]

This PR includes the following additions to SMART:

In mesh_tools:

• Add function compute_curvature, which computes the mean curvature over specified boundaries and writes out curvatures as a vertex mesh function over the parent mesh. Currently only tested in 2D.
• Add return_curvature as an argument to create_2dcell. If only a half cell geometry is desired, curvatures are computed over the whole cell contour to avoid errors at the x = 0 axis, then linearly interpolated onto the half-cell mesh.

In mesh:

• Add curvature as an attribute of parent_mesh objects (is empty by default).

In model:

• If parent_mesh.curvature is not empty, each compartment has an associated dolfin function curv_func containing curvatures from the vertex mesh function mapped onto the compartment's FunctionSpace using the new function model.mf0_to_fun.
• If user defines an initial condition that contains curv as a free symbol, then the associated curv_func is substituted into this expression when evaluating the initial condition in model.dolfin_set_function_values

In model_assembly:

• Add curv to equation variables for reaction that include curv as a free symbol.

A new version of example 2, example2_withcurv has been added to demonstrate use of the new features.

[emmetfrancis]

To do before merging:

• Test curvature calculations in 3D.
• Alter curvature definition for axisymmetric geometries defined on 2D mesh.
• Consider defining curvature as a parameter (would require implementing spatially-dependent parameters, maybe in a separate PR)

[emmetfrancis]

First two points above have been completed (curvature needed to be divided by 2 for consistency in 3D). Spatially-dependent parameters may be implemented in a future PR.

[jorgensd]

Note for later: we could use the new create_meshview function in dolfin to avoid some, But not all parallel issued

[emmetfrancis]

Added the few curvature calculation tests to the "tests" folder. The calculations in 3D are quite inaccurate here (likely very sensitive to the discretization); we could look into improving this calculation in the future.