Speed up L_custom_py

[briling]

@liam-o-marsh can be sped up. you even have an implementation of this somewhere that batches things up

Originally posted by @liam-o-marsh in #115 (comment)

[liam-o-marsh]

L_custom_c 0.06366477999836206
L_custom_py [new] 3.2500997650204226
L_custom_py [2*old+new] 581.7050670690369
L 0.061623780988156796
L_sklearn 0.4061338899191469

[briling]

#123

[briling]

closed by #123