This PR implements LoadPrevMsgMulti, which allows us to do an efficient subject-filtered walk backwards using subject tree intersection where possible.

We then use an iterative approach for finding the sourcing starting sequences where we:

1. Start off with sublists containing the subject filters and transform destinations of all of the sources
2. Walk back from the last sequence using this sublist as a filter set
3. For each source that we find the last known sourced sequence for, we can then rebuild the sublist excluding that source
4. The search for the next source gets more efficient as the sublist gets more specific

In the included benchmark, which publishes 1000 messages, sources 1 message and then publishes a further 100,000 messages, the difference is noticeable with around a 98% improvement:

Before:

BenchmarkJetStreamScanForSources/StartingSequenceForSources-24       1	   4431625 ns/op

After:

BenchmarkJetStreamScanForSources/StartingSequenceForSources-24       1      63667 ns/op

Signed-off-by: Neil Twigg neil@nats.io