Menv size controlled by inp.gpr.Mcut

[zekunlou]

To address #83.
Passed the water monomer test.

Test results for density prediction:

• Mcut: fixed (default)
  • spe H: {0: 49, 1: 49, 2: 49, 3: 49, 4: 49,}
  • spe O: {0: 31, 1: 31, 2: 31, 3: 31, 4: 31, 5: 31,}
  • problem dimensionality: 7117
  • validation: % RMSE: 4.007e-01 -.> 0.4007 %
  • test error (meV/atom): electrostatic 9.7, XC 3.0, total 0.3
• Mcut: gaussian:
  • spe H: {0: 49, 1: 47, 2: 40, 3: 31, 4: 22,}
  • spe O: {0: 31, 1: 29, 2: 25, 3: 20, 4: 14, 5: 9,}
  • problem dimensionality: 5052
  • validation: % RMSE: 4.239e-01 -> 0.4239 %
  • test error (meV/atom): electrostatic 8.2, XC 2.9, total 0.3
• Mcut: [5, 5, 4, 3, 2, 1]
  • spe H: {0: 5, 1: 5, 2: 4, 3: 3, 4: 2,}
  • spe O: {0: 5, 1: 5, 2: 4, 3: 3, 4: 2, 5: 1,}
  • problem dimensionality: 607
  • validation: % RMSE: 4.795e+00 -> 4.795 %
  • test error (meV/atom): electrostatic 43.0, XC 12.2, total 0.5
• Mcut: [1, 1, 1, 1, 1, 1]
  • spe H: {0: 1, 1: 1, 2: 1, 3: 1, 4: 1,}
  • spe O: {0: 1, 1: 1, 2: 1, 3: 1, 4: 1, 5: 1,}
  • problem dimensionality: 141
  • validation: % RMSE: 3.615e+01 -> 36.15 %
  • test error (meV/atom): electrostatic 684.8, XC 140.0, total 6.0

Possible further improvement:
Some rows are entangled with the rest with too many array reshapes, but we can in principle replace their Menv by Mcut.
I expect this come later.