Implement sparse gp variance in lammps

[YuuuXie]

Previously we have mapped variance, which only applies for power=1.

This pull request adds support for higher power by directly computing sparse GP variance in lammps. Basically, after dumping L_inv and sparse_descriptors, lammps reads the two files and in prediction, it computes kernel vector and local uncertainties.

• To use the mapped version:

  # dump coefficient file
  sgp = SGP_Wrapper(...)
  sgp.write_varmap_coefficients("varmap.txt", "me", 0)

  in lammps use command

  compute unc all flare/std/atom varmap.txt

• To use the sgp version:

  # dump coefficient file
  sgp = SGP_Wrapper(...)
  sgp.sparse_gp.write_L_inverse("L_inv.txt", "me")
  sgp.sparse_gp.write_sparse_descriptors("sparse_desc.txt", "me")

  in lammps use command

  compute unc all flare/std/atom L_inv.txt sparse_desc.txt