Inconsistent conventions for axis ordering of u_kln matrices

[mcwitt]

Sometimes our usage aligns with the pymbar convention where the first two dimensions correspond to sampling and evaluation state respectively, e.g.: https://github.com/proteneer/timemachine/blob/c4369bdd3accbb88d48cabd24c9dc209d19d94f6/timemachine/maps/estimators.py#L54-

and sometimes we use a convention where the meaning of k and l is swapped, e.g.:

timemachine/timemachine/fe/energy_decomposition.py

Lines 82 to 88 in c4369bd

	u_kln_by_component = np.zeros((n_components, K, K, n_frames))
	for comp in range(n_components):
	for k in range(K):
	u_fxn = states[k].batch_u_fns[comp]
	for l in range(K):
	xs, boxes = states[l].frames, states[l].boxes
	u_kln_by_component[comp, k, l] = u_fxn(xs, boxes)

Given that inconsistencies can easily lead to bugs, it would be useful to align on a single convention.