Radios translations: cuda backend creates way too many intermediate arrays

When using translation_movements_files, it seems that one intermediate cuda array is created for each angle.

Processing sub-volume ((0, 10001), (445, 675), (0, 2560))
Allocate (1, 230, 2560) float32
Allocate (1, 230, 2560) float32
Allocate (1,) int32
Allocate (10001, 2) int32
Allocate (10001, 2) float32
Reading data
Read subvolume (10001, 230, 2560) in 63.99 s
Applying flat-field
Dumping data to /data/scisofttmp/paleo/debug/PP/rec/flatfield_Cell1_LPSC_NMA_Pristine_uCT_0001_rec/flatfield_Cell1_LPSC_NMA_Pristine_uCT_0001_rec_00448.hdf5
Taking logarithm
Applying radios movements
Allocate (229, 2560) float32
Allocate (229, 2560) float32
Allocate (229, 2560) float32
Allocate (229, 2560) float32
Allocate (229, 2560) float32
... # 10,000 times !