5 and 17 are culprits.  After the fast FiniteDifference is moved from pysindy
to derivative, 5 will be faster.

Also gitignore output files from 17.

We don't need to test legacy ensembling behavior for every optimizer,
since the ensembling part doesn't depend upon the inner optimizer.  We also
don't need 10x10 bagging x library_bagging, we just need 2x2 to test
functionality.

Old test durations:
4.82s call     test/optimizers/test_optimizers.py::test_legacy_ensemble_odes[StableLinearSR3]
1.69s call     test/optimizers/test_optimizers.py::test_legacy_ensemble_odes[MIOSR]
0.36s call     test/optimizers/test_optimizers.py::test_legacy_ensemble_odes[TrappingSR3]
0.34s call     test/optimizers/test_optimizers.py::test_legacy_ensemble_odes[FROLS]
0.11s call     test/optimizers/test_optimizers.py::test_legacy_ensemble_odes[SSR]
0.04s call     test/optimizers/test_optimizers.py::test_legacy_ensemble_odes[ConstrainedSR3]
0.04s call     test/optimizers/test_optimizers.py::test_legacy_ensemble_odes[SR3]
0.04s call     test/optimizers/test_optimizers.py::test_legacy_ensemble_odes[STLSQ]

New test duration:
0.02s call     test/optimizers/test_optimizers.py::test_legacy_ensemble_odes

Also, make test/differentiation/test_differentiation_methods.py just
test/test_differentiation.py.  Unneccessary nesting.

Previously, test_nan_derivatives asked whether the model fit by SINDy when
FiniteDifference gave NaN endpointswas close to the model fit without nans.
This is an integration test that is really asking whether FiniteDifference is
calculating derivatives in a consistent with drop_endpoints=True and False.

Also, a bit faster but this test was fast to begin with.

Previously, test_complexity and test_coefficients just tested that
- required names are present in SINDy object
- when fit on some amount of data, less than 10 terms were nonzero
- These are essentially the same test, since model.complexity is just
  a wrapper around it's optimizer's complexity property

But Lorenz has 7 nonzero terms, and when I changed the data to speed up
tests, data had a few more nonzeros.

We don't need 500 timepoints; shortening to 50.

Also, data_sindypi_library and data_pde_library had an implicit
dependence upon data_lorenz; this commit makes that dependence explicit.

A lot of test files were nested under a directory with just one file.  It's now
one less click to view test results or files.

see commit 3445cb9