calculate_G4 and cutoff functions

[andrew_peterson]

I think that in the Behler fingerprint scheme, calculate_G4 is incorrect. According to Artirth:

http://pubs.acs.org/doi/suppl/10.1021/nl5005674/suppl_file/nl5005674_si_001.pdf

each term should be multiplied by the product of three cutoff functions. Our code looks like it is multiplying by the sum of the three cutoff functions.

I noticed a discontinuity when plotting an angular fingerprint term versus bond length (abruptly goes to zero as the bond length exceeds the cutoff threshold), and think this explains it.

[andrew_peterson]

This became a problem here:

https://bitbucket.org/andrewpeterson/neural/commits/a4b518f776826611a6096dd3a7db2ca409fc65b7

[Alireza Khorshidi]

This is now restored to original format of Behler.