Fix thread-safety issues in MuonSegment

There's no need for MuonSegment to cache a vector of pointers to RIO_OnTrack objects, essentially as a convenience. If clients need them, they can be built on the fly on demand: alternatively, the rioOnTrack(int) method can be used to loop through them. This does require adding some protection in case of the presence of competing ROTs, though.

Also, there turns out to be no reason at all for the m_associatedSurface member to be mutable.