Sparse + Parallel

[rtbs-dev]

Updates the core associations.py to use pydata/sparse as a backend

• add flat2sq and sq2flat as sparse squareform re-writes
• dispatch norm_diag to preserve sparsity in sparse grammian norms (e.g. cosine-sim)
• parallelize core forest_pursuit loop with joblib

All together, we can do precision matrix/edge support reconstruction on sparse binary input with very little memory change from the pre-existing data.

10k variables and 50k observations can be made into a dependency graph estimate in under 10 sec on a laptop.

[rtbs-dev]

Unit tests still failing for strange shapes, but otherwise working for forest_pursuit_cts.