Fix small cutoff neighbor

[chc273]

Summary

In the current cell list implementation for finding neighbors, the space is divided into ~O((cell length / cutoff)^3) cubes for fast neighbor finding. This causes issues when the user chooses unrealistic small or even zero cutoffs.

In this PR, I am doing the following:

1. For reasonable cutoff (default to > 1.0 Angstrom), I am calculating the neighbors directly using the algorithm.
2. For small cutoffs (e.g., < 1.0 Angstrom, including neg cutoffs), I am calculating the 1.0-cutoff neighbor list, and use a boolean mask to keep those neighbors having distances less than the cutoff afterward. This will also make sure that if the users input negative cutoff, the returned neighbors will be empty.

[coveralls]

Coverage decreased (-0.6%) to 83.101% when pulling 00ecc46 on chc273:fix_small_cutoff_neighbor into 71c8ab8 on materialsproject:master.