Remove CudaChunkedPipelineLimitedMemory class

For the record, nabu comes with two (full-field) pipelines:

1. "chunked": load several lines of all the projections (a "chunk" of radios): ChunkedPipeline
2. "grouped": load all lines from a group of projections (a "group" of radios): GroupedPipeline

Pipeline (1) is more efficient as it only loads the data needed for reconstruction. The memory transfers are kept minimal.

Pipeline (2) is adapted for the following use cases:

• (a) Datasets with a very large number of projections (such that n_projections * n_slices_to_reconstruct * slice_width does not fit in memory)
• (b) Steps that need to load most-or-all the projection image: CTF, rotation with a large angle, Paganin with large kernel footprint

There is currently a CudaChunkedPipelineLimitedMemory which tries to address use case (a) while still retaining the "chunked" approach. It was kept for legacy, but the implementation is really cumbersome (both the class and the branching in ChunkedReconstructor). It would be good to get rid of this class altogether.

The current implementation of GroupedPipeline writes temporary results (sinograms) to a file. But in some cases n_projections * n_slices_to_reconstruct * slice_width does fit in host memory (although not in GPU memory), so the sinograms can simply be retained as an array.

So to summarize:

• Make GroupedPipeline able to keep the sinogram as an array instead of writing to a file
• Ensure GroupedPipeline can replace the precise use case of CudaChunkedPipelineLimitedMemory
• Remove CudaChunkedPipelineLimitedMemory and related code
• End to end tests