Performance improvement for cbm::algo::tof::Hitfind::CalibRawDigis().

A fix for the mysteriously slow calibration of TOF digis observed at the DC.

The origin was something even way dumber than than the possible cache-issues which were discussed:

In the calibration loop, a hard copy of the walk coefficients for the given channel was carried out in each iteration. This slipped in while I implemented the steering layer last week. As a fix, references are used instead.

This code is way faster. Total execution was cut by roughly a factor of four, but this still includes some overhead, so the true speedup is likely higher.

As a side note, this emphasizes the point that we are bottle-necked by calibration here, not by actual cluster building. The benchmarks I showed at the collaboration meeting were for the hit-finding step only, and remain valid. We hence should consider doing calibration in a more clever (parallel) way.

May be of interest to: @f.uhlig @fweig @j.decuveland @s.zharko @se.gorbunov @v.friese @p.-a.loizeau

Will set to auto merge.