Add a Hessian test to PoissonLogLikelihoodWithLinearModelForMeanAndProjData (and fix Hessian)

[robbietuk]

Addresses #911

• Adds test_objective_function_Hessian which utilises objective_function.accumulate_Hessian_times_input to test the concavity condition x^T H x <= 0.

• Modifies actual_accumulate_sub_Hessian_times_input_without_penalty (and approximate version) in PoissonLogLikelihoodWithLinearModelForMeanAndProjData to be a concave functions (plus -> minus)

[robbietuk]

I cannot workout why this was closed, it is a fix to the actual_accumulate_sub_Hessian_times_input_without_penalty

[robbietuk]

accumulate_sub_Hessian_times_input is used in other places (e.g. spatially variant penalty strength) that need to correct for the new negativitity.

For Poisson methods: add_multiplication_with_approximate_sub_Hessian_without_penalty needs to be negative too. This will break backwards compatability and will need to be documentented as well as uses of the method adjusted to accommodate.

Uses of accumulate_sub_Hessian_times_input include:

• SqrtHessianRowSum->compute_Hessian_row_sum

Uses of add_multiplication_with_approximate_sub_Hessian_without_penalty include:

• SqrtHessianRowSum->compute_approximate_Hessian_row_sum()
• `OSSPSReconstruction->precompute_denominator_of_conditioner_without_penalty()
• Any python code (none within STIR project - therefore needs to be well documented)

Of the SqrtHessianRowSum and OSSPSReconstruction uses, the output should have positive voxel values (these changes result in negative voxel values given a positive input)