perf(sync): benchmark and optimize scoped first-view sync paths

[marcus-pousette]

What This Adds

• product-shaped scoped sync benchmarks for opening a node on a fresh client: fetch the first page of a node's immediate children, with and without payloads
• matching balanced subtree re-sync benchmarks for the case where a restarted client already has that scope locally and needs a normal non-empty scoped reconcile
• a --first-view metric so sync time and first local read time are measured separately
• sync-server fixture priming and reuse, so one-time upload/setup cost is measured separately from steady-state open cost
• the protocol and Postgres sync-path changes needed to make these scoped paths fast enough to be meaningful

Benchmark Results

Reference first-view case: open a known node, fetch its first 10 immediate children, and include 512B payloads.

Fresh First View: Before vs After

• before, on the old local Postgres 10k open path:
  • structure-only: about 18.06s total, with about 17.91s in server listOpRefs(children)
  • payload variant: about 26.77s total, with about 26.55s in server listOpRefs(children)
  • local SQLite apply and first local read were only a few milliseconds, so the bottleneck was the server-side scoped query path
• after, in this PR:
  • local Postgres sync server, 10k tree: 27.5ms total time to first view, 27.1ms sync, 0.38ms first local read
  • local Postgres sync server, 100k tree: 27.6ms total time to first view, 27.2ms sync, 0.50ms first local read
  • remote sync server, 100k tree: 227.2ms total time to first view, 226.9ms sync, 0.79ms first local read

Same Subtree Re-Sync On A Non-Empty Client

• local Postgres sync server, structure-only scoped re-sync:
  • 10k: 41.0ms sync-only
  • 100k: 59.2ms sync-only
• local Postgres sync server, payload scoped re-sync:
  • 10k: 38.2ms sync-only
  • 100k: 56.5ms sync-only

What These Numbers Cover

• the fresh first-view numbers are the empty-local-result case, where the protocol can sometimes skip a full RIBLT round and direct-send a tiny scoped result
• the re-sync numbers are the non-empty case, so they measure the normal scoped reconcile path instead

Setup Cost Measured Separately

• remote sync-server fixture prime for the same 100k first-view case is about 108.2s, and this PR measures that separately so upload/setup does not get mixed into the open-path numbers

Scope

This PR is about scoped first-view sync and subtree re-sync performance.

The later large-doc hot-write benchmark suite and its setup optimizations live in #113.