Liz speedup cat

[ElizabethSantorellaQC]

See #3

Split categorical-dense is now competitive with split0.1 and sparse, but not always fastest. There is more to do, but no harm in merging this. I've parallelized stuff, so I expect a greater speedup on a machine with more cores.

More performance benchmarks available in the README here: https://github.com/Quantco/quantcore.matrix/tree/benchmark

                                                                   library_name   storage  threads  single_precision  n_iter  runtime  offset  intercept  num_nonzero_coef  obj_val  rel_obj_val
problem_name                    num_rows regularization_strength                                                                                                                                
wide-insurance-weights-l2-gamma 24944    0.001                    quantcore-glm     dense        8             False       8   0.4407   False     6.9434               273   8.1218   0.0000e+00
                                         0.001                    quantcore-glm  split0.1        8             False       8   0.1080   False     6.9434               273   8.1218   0.0000e+00
                                         0.001                    quantcore-glm       cat        8             False       8   0.1376   False     6.9434               273   8.1218   1.7764e-15
                                         0.001                    quantcore-glm    sparse        8             False       8   0.1653   False     6.9434               273   8.1218   1.7764e-15

[MarcAntoineSchmidtQC]

REALLY exciting.

[ElizabethSantorellaQC]

Needs to be merged first: Quantco/glum#264

[tbenthompson]

Separate review of the C++ parts: Looks great!