Improve consistency in parameter names

[bellet]

It would be nice to follow a consistent naming convention for parameters and be as consistent as possible with sklearn. For instance:

• In supervised versions of weakly supervised algorithms, num_constraints should be renamed n_constraints, num_chunks to n_chunks
• In LMNN, the parameter k could be renamed n_neighbors like in sklearn's KNeighborsClassifier
• There is also tol and convergence_threshold which are both used to refer to optimization tolerance (we should always use tol which is quite standard, cf `scipy.optimize)