Some ideas

[clbarnes]

• Use PEP517 build spec to handle numpy/cython build-time dependencies
• I think the vertex similarity function is commutative, so there's a ~2x speedup to be had by only computing one triangle of the output
• Unit test to verify that the same results are produced
• More flexible/aggressive coercion of input type (also removes pandas dependency)