Task #5844: COBALT PVSS: zero dynarrays pre-obs (while also logging obs ID, data product names, etc). This is nice to have, because some events are written conditionally. If a subband has no dropped data, it's better to init that value to 0.0f. Some minor cleanups related to PVSS event logging is also included.

Task #5841: Moved LTA feedback functionality from Parset to LTAFeedback class. Added createFeedback tool to (re)create LTA feedback files, and removed said functionality from createHeaders.

Task #5926: Failed write() is fatal in non-real-time mode, and fixed skipping over missing data in case of write() failure.

Task #5830: revert most KVpair changes. The extra overloads clash on either 32 or 64 bit. Adding overloads will also fail, as they cannot be mapped to PVSS data types in the PVSSGateway (default case there is to try string). So overloads -> template won't work either. Agreed with Ruud to also remove the vector<int> support (already not mapped). Then to solve the original data type problem, adapt a few logged data types from uint or ulong to int. While at it, adapt a few to bool from string ("0" or "1").

Task #5830: address review comments. Improve data point logging in CorrelatorPipeline. Move INIT log earlier in rtcp.

Task #5833 (in branch of Task #5830): move writing of dropped PVSS data point to a less frequently executed place

Task #5830: outputProc: cast PVSS data point vals for 'written' and 'dropped' to float before sending to conform to PVSS data point defs.

Task #5830: write CobaltOutput data points using the MACIO RTmetadata interface. Perhaps, the 'dropping' data point is written too late. Also adapt up 2 test cases + createHeaders.cc with RTmetadata dummies that won't try to log/send anything.

Task #4818: Catch feedback/augment exceptions in the OutputThread, to make GPUProcIO code a bit cleaner

Task #4625: Forward LTA feedback from OutputProc to GPUProc using the control stream (instead of the logs), and write the LTA feedback to disk in rtcp.cc

Task #4643: use THROW_SYSCALL() macro instead of 'throw SystemCallException(...)' to ensure that errno is loaded immediately and not after it was possible clobbered by string() or Backtrace(). Replace open-coded version in a few instances where the problem had already bitten. Note that all pthread_*() do not set errno, but return it (or 0), so keep those as-is. Fix up errno passing of pthread_kill() in this regard.