ITS GPU Seeding vertexer: use warp reductions for fit peak kernel

[cima22]

ITS GPU seeding vertexer: fit vertex candidates with one warp per candidate

fitPeaks fitted each vertex candidate with a single thread, looping serially over every line in the candidate's z-window. A TF has few candidates but windows of up to thousands of lines, so a handful of threads with large windows set the kernel duration while most of the GPU sat idle. The kernel now assigns one warp per candidate: the lanes stride over the window, and the normal-equation sums, time bounds, contributor counts and residuals are combined with warp shuffle reductions, so every lane holds the result.

• fitPeaksKernel replaces the thread-per-candidate kernel.

• GPUCommonAlgorithm.h: adds warp_reduce_add/min/max<Width>() (XOR butterfly, result in every lane) next to warp_broadcast

• ClusterLines.h: the normal equations, the solve and the fit results move into a LineFit base class, shared by the CPU ClusterLines (sequential) and a new device-only WarpFit.

The handler aborts if the device's warp size differs from the compiled gpu::WarpSize.
On 4 Pb-Pb raw TFs (run 560161, AMD MI50): fitPeaks kernel time summed over the TFs went from 80.2 ms to 3.5 ms (23×), and the GPU seeding vertexer is 10–13% faster per TF. Vertices and tracks are identical to before on both GPU and CPU.

[alibuild]

Error while checking build/O2/fullCI_slc9 for a7472dc at 2026-10-14 20:02:

No log files found

Full log here.