chain/exchange BlockSync port: snapshot-free full-node boot

[Reiers]

Part of #87 (Full Node epic). Stage A — the snapshot-free sync engine.

Port Lotus chain/exchange (BlockSync, pure-Go libp2p protocol) so a Lantern full node boots without a snapshot:

• Anchor at a recent F3-finalized tipset (multi-source quorum — already built).
• BlockSync backward to backfill headers/messages the operator wants.
• Gossip for the head.
• Never import 76 GB, never re-exec from genesis.

Security angle (the pitch): a Lotus snapshot import trusts a single snapshot host for all pre-snapshot state. Lantern's anchor is a multi-source, F3-finalized, BLS-verified quorum — strictly stronger boot trust. "Full node, no snapshot, minutes not days."

[Reiers]

Scope refinement after a fresh lotus master survey:

• Protocol: /fil/chain/xchg/0.0.1. Request{Head []cid, Length ≤ 900, Options bitfield} where Headers and Messages are independent option bits — headers-only deep-tail fetch is a first-class protocol option. That matches our retention shape exactly: headers-only for the deep tail (~380 MB for all mainnet headers), Headers|Messages only for recent finalities.
• Response semantics include Partial (101) — a peer may legally return fewer tipsets than asked; client must loop.
• Lotus's whole chain/exchange package is ~2.1k LOC including the server; a client-only lift is well bounded (~1.2k). net/chainxchg (170 LOC) is our starting stub.
• Peer management: lotus tags good exchange peers in the connmgr (SuccessPeerTagValue) + shuffles a preferred-peer prefix. Reuse our existing peer plumbing + trusted floor rather than lifting their peer_tracker wholesale.
• Serve-side (being a BlockSync server) is a separate follow-up: it strengthens the swarm story but is not needed for snapshot-free boot.

Order: client first, wired behind the Full tier; server later.

[Reiers]

Stage A (client) implemented — PR #110.

The requester half of `/fil/chain/xchg/0.0.1` is done and green. A bridge-off node (`--no-fallback-rpc`, #76) can now backfill a gossip head+N gap by pulling the whole CID-verified header chain from a peer in one request, with no Glif and no gateway.

• Headers-only requests, parent-ward walk, handwritten CBOR cross-checked vs Lotus `chain/exchange`.
• Verification chains the CIDs (level 0 == requested head, each level == prior level's Parents), so a malicious peer can only refuse, never splice us onto a fake chain.
• Partial(101) tolerated; NotFound/GoAway rotate peers; trusted-floor-first ordering; bounded reads.
• Wired into blockingest as the preferred parent-walk gap resolver (per-CID bitswap walk stays as fallback).
• Tests: happy path / partial / NotFound / spliced-chain-rejected / no-peers. build + vet + fmt + test all clean.

Remaining for the epic: daemon wiring (use the client for the anchor seed → also closes #109), and the server side (serve real tipsets instead of NotFound-only) as a separate follow-up.