Vectorize atom index computation in L-BFGS optimizer

[falletta]

Replace Python for-loops over systems with vectorized tensor operations in the L-BFGS optimizer's padding/unpadding helpers and mask construction.

[abhijeetgangan]

Could you run #365 (comment) and see the performance did not regress?

[falletta]

Sure, I’ve done two consecutive runs of the script in #365 with and without these changes, and obtained:

Before:  L-BFGS       178.79       1000/1000          0.0001
After:   L-BFGS       165.76       1000/1000          0.0001

A second pair of consecutive runs gives:

Before:  L-BFGS       185.18       1000/1000          0.0001
After:   L-BFGS       171.04       1000/1000          0.0001

So it looks like, besides timing fluctuations, there might also be some minor optimization.