Add eigenstrapping null generation

[VictorBarnes]

• Created new module nulls.py with a main function eigenstrap to generate eigenstrapped nulls. This function uses _eigenstrap_single to generate each null.
• Added eigenstrap as a class method to EigenSolver
• Added get_eigengroup_inds to eigen.py
• Removed joblib as optional dependency and made it a project dependency
  • Removed the check for joblib in io.py and removed corresponding test in test_io.py
• Added tests in test_nulls.py
• Added a tutorial in eigenstrapping.ipynb

---

Commit f6f95f1 contains code to exactly match the nulls generated in the original eigenstrapping code by Nikitas Koussis if we need to go back to compare. The key differences include:

• Set the default eigendecomposition method to project the data into mass-orthonormal space (as discussed above). See also this lapy issue. Regress is also included as a decomposition method and is a good option to use when the input data contains NaNs. The mass matrix is an optional input accordingly.
• Consolidated add_res and permute into one argument residual which can either be None, add or permute
• Use all modes in computation (including constant mode). This is to be consistent with the rest of the toolbox where all main functions always take in all modes as an input. This also makes eigenstrap cleaner. To avoid division by zero in the normalisation, we set the first eigenvalue to be 1. The first mode doesn't get rotated but keeping this in will preserve the mean of the nulls when resample=None.
• Require all modes, data, and mass matrix to already be masked when passed in. Again, this keeps this cleaner and also avoids conflicts with the EigenSolver which already does masking. Since eigenstrap is a method of that class, adding a masking option in eigenstrap would lead to conflicts with the EigenSolver mask.
• Removed normalize as an option.
• Seeding is handled in two stages to ensure reproducibility when parallelising. First seed is used to initialize a master random number generator (RNG) that generates an integer seed for each null map. Then each null uses its allocated integer to generate its own RNG to use for all rotations/permutations of that null. This ensures that each null is independent of the number of parallel jobs used.
• Refactor resample to include multiple options: 'match' to match the sorted of the original data, 'zscore' to z-score and rescale to original mean std, 'mean' to preserve the mean, and 'range' to preserve the minimum and maximum. Default is None for no resampling.

[magnesium2400]

why doesnt the nulls sphere PSD match the original any more? test_psd_preservation still passes i presume

[VictorBarnes]

why doesnt the nulls sphere PSD match the original any more? test_psd_preservation still passes i presume

Looks like I just forgot to rerun that notebook before committing. The PSD looks like that when resample="range". Whoever makes the next commit can just make sure to rerun it.