Priortise FastRetrieval candidates from indexer

[rvagg]

From #35 (comment)

This has been discussed in the context of autoretrieve for some time now but I think originally they were all set to true but now it appears Boost has this wired up so keep-unsealed-copy config = FastRetrieval metadata.

How:

1. The Metadata field from the indexer results is base64 (padded) encoded varint-prefixed dag-cbor which can be decoded via https://github.com/ipni/index-provider/blob/19931fd5c692e5efd38093c1764cf0a3ca464af4/metadata/graphsync_filecoinv1.go#L66
2. FastRetrieval = true entries should get priority in the queryCompare function for the prioritywaitqueue when we have multiple queries returned and waiting for attempts.

Questions:

1. Should prioritisation be even higher than this? If candidate A is FastRetrieval=false and B is FastRetrieval=true and A returns its query first, should we hold off on attempting a retrieval and give B a chance to come back first?
2. Perhaps we should shorten the first-byte timeout for FastRetrieval=false candidates under the assumption that they may be attempting an unseal?
3. If we have >X candidates and some high percentage of them are FastRetrieval=true, perhaps we should filter out the others and only attempt the ones we assume have unsealed copies? e.g. 10 candidates for a CID, 7 of them are FastRetrieval=false, don't even bother attempting the other 3?

[hannahhoward]

Let's talk through this in more detail. As we start thinking about how we get the data the fastest, here are some thoughts:

• I think we need to find a way to skip the query phase. My proposal is as follows:
  • simply propose free deals to each returned peer, one by one -- they'll respond if they don't accept free
  • prioritize the order by:
    • first, whether peers we have an existing libp2p connection to first (since this is a ~0.5s penalty)
    • second, by peers who advertise "fast retrieval" and/or "verified data" first (since these are most likely to return free retrievals)
    • eventually, by tracking information about peers and prioritizing the most efficient ones (probabilisticly)