Fix subnet samples (posterior_scale.T issue)

[elcorto]

Closes #274

This fixes the posterior_scale.T issue discusssed in #274 (FullSubnetLaplace.sample()) and adds tests to make sure FullSubnetLaplace.sample() and DiagSubnetLaplace.sample() do the right thing.

We also add cov sampling tests for all other Laplace variants in tests/test_baselaplace.py and tests/test_lllaplace.py.

[elcorto]

I can reproduce the pytest-with-asdfghjkl fail locally when I install git+https://git@github.com/wiseodd/asdl@asdfghjkl as the action does. Then the max error is ~0.7 instead of the allowed 0.01:

tests/test_baselaplace.py::test_laplace_functionality[LowRankLaplace-classification] >>>>> tensor(0.7377)
FAILED
tests/test_baselaplace.py::test_laplace_functionality[LowRankLaplace-regression] >>>>> tensor(0.7272)
FAILED

but only the combo LowRankLaplace + the above asdfghjkl lib. Maybe one can xfail that by testing for LowRankLaplace and a specific version (range) of the library? At least that's a quick solution to make the (PR) tests pass. Solving the asdfghjkl issue is a separate topic I guess.

[elcorto]

The lint fail

93 |     assert res != math.nan
   |                   ^^^^^^^^
94 |     assert isinstance(res, float)
95 |     assert res > 0
   |

PLW0177 Comparing against a NaN value; use `math.isnan` instead

is not due to changes in this PR. Also, I'm with ruff here :)