kyleaoman/martini: v2.0.3

This release adds significant changes to the SPH kernel classes.

Instead of trying to walk users through the best practice for setting up an "adaptive kernel" that will fall back to other kernels when the approximations used for the kernel line integrals lose accuracy, the best practice is now implemented in a set of recommended kernels, with the individual underlying kernels now moved to private members shadowing the recommended "adaptive" versions. So for example the old WendlandC2Kernel becomes _WendlandC2Kernel, and is replaced by a WendlandC2Kernel that is adaptive and falls back to the DiracDeltaKernel or GaussianKernel when needed to achieve the target accuracy. The new approach is reflected in updated example notebooks and a new narrative docs page for SPH kernels.