Remove use of operations with different results on AMD and Intel

Currently, the AMD machines are not used for testing as they produce slightly different results.

This MR removes two operations which yield different results for AMD and Intel.

• PrVeloUT uses rsqrt in some calculations, replaced it with 1./sqrt.
• PrForwardTracking uses rcp a less precise division, replaced it with a simple division.

Both changes change the references. A very small loss in throughput is expected.

With the changes above, I could produce identical log files for the tests hlt1_reco_baseline and hlt2_reco_baseline_with_mcchecking.

Note: Tested with v2+dbg and v3+opt compiler flags, i.e. SSE and avx2 were tested.

Reference updates in Moore!916 (merged) and Brunel!1156 (merged)