Force constant model speedup

Description

Several optimizations related to cluster space building and in particular enforcing of rotational sum rules.

Summary of changes

• hiphive/core/atoms.py
  • spos_to_atom: sped up arithmetic by using NumPy arrays instead of python loop
• hiphive/core/rotational_constraints.py
  • _create_Huang_constraint and _create_Born_Huang_constraint: sped up matrix construction by using NumPy arrays and SciPy sparse matrices instead of SymPy sparse matrices
  • enforce_rotational_sum_rules: removed the conversion from SymPy matrices to SciPy matrices since the constraint functions now return SciPy matrices directly
• hiphive/core/utilities.py
  • nullspace: sped up construction of nullspace vectors by initializing vectors as empty SparseMatrix instead of converting them to SparseMatrix after building them as lists of lists. The calls to _new are now faster
• hiphive/force_constant_model.py
  • get_fcs_sensing: added argument option to return sensing matrix as sparse (SciPy coo_matrix). Default option is False, so existing code using this function should not be affected
• hiphive/force_constants.py
  • _sanity_check_dict: sped up sanity check by using a NumPy array instead of a python loop
• hiphive/utilities.py
  • extract_parameters: added an option to use a sparse iterative least squares solver, which reduces runtime for large systems where NumPy's linalg.lstsq becomes inefficient