Multi-threaded material budget LUT creation

[trwenz]

Introduces multi-threaded creation of the material budget lookup table, reducing it from hours to minutes. Creating the LUT walks every cell of every layer through TGeo and the cells are independent, so they are spread over TBB tasks with one TGeoNavigator per thread. The innermost 20 layers at 60 trials/cell drop from 28 min to 72 s on 28 cores.

Effective only with ROOT >= v6-36-10-alice3, which removes a per-query thread-id lookup and the false sharing between per-thread scratch buffers of TGeo shapes. On older ROOT the parallel path is correct, just slower -- it saturates near 12x.

All layers map onto a single flat cell index so the load stays balanced despite very different cell counts per layer; a binary search maps a flat index back to (layer, iz, iphi). The worker navigators are given back at the end.

meanMaterialBudget() takes an optional navigator: a caller passing its own runs lock-free, a caller passing none shares gGeoManager's and still takes the mutex. Deciding from the argument keeps it local, so process-global state cannot break it.

Thread count comes from the new populateFromTGeo() argument, falling back to NTHREADS_MATBUD; the default is the previous serial path. Results are independent of the thread count -- compareMatBudLUT.C checks a parallel LUT against a serial one cell by cell, and they match exactly over all 129523 cells.

Supervised-by: Sandro Wenzel sandro.wenzel@cern.ch

[trwenz]

This plot shows the measured Speedup of the parallel implementation using the old ROOT version and the patched ROOT version.

[shahor02]

Excellent, thanks!