Accelerate ThreePIE engine

[daurer]

Starting point is existing implementation of CPU version - tested on real data.

[ltang320]

- Task:

Implement ThreePIE_cupy that mirrors the CPU ThreePIE and uses the same GPU infrastructure and kernels that we already have (EPIE_cupy) in stochastic.py.

- Implementation road-map:

• Propagate through the slices → kernel 1: multislice_fw
• Copy last slice to ptypy → kernel 2 : last_slice_copy_to_ptypy
• Fourier update (from EPIE_serial)
• From ptypy copy the last slice out → kernel 3: ptypy_copy_to_last_slice
• Back propagate through the slices → kernel 4: multislice_bw

[ltang320]

Migrate the multi_slice_epie CPU engine to GPU engine:

1. Rewrite it in serial version (from views operation to array operation)
2. Compare the reconstruction results CPU multislice version vs. serial version
3. Migrate it to GPU final....ly