CudaFullFieldPipelineLimitedMemory: crash when reconstructing first slice with a large margin

Reported by @majkut

Reconstructing a dataset with a margin of 71 pixels entails to processing at least 143 lines of each radio.
If chunk_size = 130 (i.e at most 130 radios lines fit in device memory), then the class CudaFullFieldPipelineLimitedMemory is used. However, reconstructing the first slice only requires to load 72 lines, so for the first slices we have chunk_size > delta_z which fails the assertion of CudaFullFieldPipelineLimitedMemory.

Possible solutions:

1. In LocalReconstruction: use CudaFullFieldPipeline for the first slices, and CudaFullFieldPipelineLimitedMemory for the others
2. Lift the limitation delta_z > chunk_size in CudaFullFieldPipelineLimitedMemory

(1) is cumbersome to implement. Moreover, it seems that removing the (old) assertion delta_z > chunk_size works with minimal modification.