listmode projectors/ parallelproj for listmode

[KrisThielemans]

Current listmode objective functions are essentially written in terms of ProjMatrixByBin. This is not suitable for GPU.

Some pointers:

• actual multi-threaded gradient calculation is here
  

  STIR/src/recon_buildblock/distributable.cxx

  Lines 599 to 607 in 8ced2d7

  	LM_distributable_computation(const shared_ptr<ProjMatrixByBin> PM_sptr,
  	const shared_ptr<ProjDataInfo>& proj_data_info_sptr,
  	DiscretisedDensity<3, float>* output_image_ptr,
  	const DiscretisedDensity<3, float>* input_image_ptr,
  	const std::vector<BinAndCorr>& record_ptr,
  	const int subset_num,
  	const int num_subsets,
  	const bool has_add,
  	const bool accumulate)

• This takes BinAndCorr:
  

  STIR/src/include/stir/Bin.h

  Lines 110 to 114 in 8ced2d7

  	struct BinAndCorr
  	{
  	Bin my_bin;
  	float my_corr;
  	};

  
  which sadly is not a Bin (although that has a value field, and in the caching we re-use that for the "additive correction" I believe)
• non-multi-threaded calculation of gradient, i.e. without caching is here
• the Listmode objective hierarchy is split into 2 (not always logically)
  • PoissonLogLikelihoodWithLinearModelForMeanAndListModeData
  • PoissonLogLikelihoodWithLinearModelForMeanAndListModeDataWithProjMatrixByBin, adding stuff for computing sensitivity and obviously the gradient

Probably with not too much work the PoissonLogLikelihoodWithLinearModelForMeanAndListModeDataWithProjMatrixByBin can be generalised. We might need (and seem relatively easily add)

• virtual void ForwardProjectorByBin::forward_project(std::vector<Bin>&), storing results in the value. (Currently a by bin version is disabled).
• virtual void BackProjectorByBin::back_project(std::vector<Bin>&)

These would have to be implemented for the ProjMatrixByBin and Parallelproj projectors, but throw an error by default (for other projectors).