Draft: Use a single configured sequence instance

This MR highlights the possibility to use a single instance of the configured_sequence_t type. This would facilitate the monitoring effort (!641 (closed)), because aggregation would no longer be needed.

All the algorithms are re-entrant, which made this possible.

It turned out to be easier to have multiple instances (one per thread/stream) of the host and device memory managers and the argument manager instead of passing a single instance of the sequence to multiple instances of the stream. Having a single instance of the configured_sequence_t would also be possible, but require a bit more refactoring.

The name Stream is perhaps not appropriate anymore.

For your consideration: @dcraik, @graven, @dovombru, @dcampora