GPU split kernels

The split kernel pair against the fused kernel it is measured against

4/4 passed in 11ms

Test	What it does	Result
test_split_matches_arbiter	The split pair's variance and xi match the f64 arbiter at VAR_TOL and XI_TOL.	✅ pass
test_split_matches_fused	Storing the phase rather than re-deriving it gives the fused kernel's slope, variance and pilot SNR within the f32 tolerances: the two wrap modulo 2*pi at different points, so a wider gap means the split arm times other physics.	✅ pass
test_split_reduces_the_same	The split pair uses the same block grid and workgroup tree as the fused one.	✅ pass
test_split_times_two_passes	Both split passes report their own kernel time, inside the wall clock.	✅ pass