- split caching in separate functions for reading a batch from
list-mode file or cache
- always use LM_distributable_computation for gradient

Resulting code avoids duplication and is faster when caching is not used.

The code pre-allocated bins for every thread, but local variables
are allocated on the stack, which should take no time. Removing
th epre-allocations simplifies the code, and prepares us for re-use
for Hessian calculation etc

- make list-mode loop into templated function
- introduce separate functions for gradient and Hessian

- move LM_distributable_function implementation to .txx as it now uses
 a template for the callback
- implement actual_accumulate_Hessian... for LM
- try to implement accumulation of log-likelihood

WARNING: LM_distributable_computation is now generic and
incompatible with the previous one (which was only for the gradient)

Hessian function and log-likelihood value are currently untested/

Also expand test to use all subsets, make it slower of course.

test_PoissonLogLikelihoodWithLinearModelForMeanAndProjData and test_priors
contained nearly similar code, so moved it to a new class.

This needed minor modifications to the tests, but also addition of
GeneralisedObjectiveFunction::compute_value()

- compute_objective_function() is now implemented as well
- Tests are (almost) a duplicate of the ProjData version, but
  gradient-test is disabled as for this sample data it is too slow

Also update to upload-artefacts@v4

also minor clean-up

Cannot figure out what the problem is, but the test works on Ubuntu and Windows