perf: v0.8.0 benchmark gaps against the phase-4 targets

[xsmyile]

The final benchmark before v0.8.0 (v0.7.0 against 4cc0970: native, 8 SMTP connections, v0.7.0 and HEAD runs interleaved, n = 6) left these gaps against the phase-4 targets.

• S2 (mail with a 150 KB attachment): 1.27×, target 1.39×. 946 → 1 204 stored/s. Against phase-3 master (4289944) it is only 1.14–1.16×. CPU during S2 is 108–119 % because MIME parsing is serial; the lever is perf(server): parse MIME in parallel with the write #117.
• S1 and MIX are about 12 % slower than phase-3 master, at 0.87–0.93× (target ≥ 0.95×). Bisecting points to the phase-4 merge 22d54e3; PRs fix(tui): ignore duplicate live message:new for an already-listed id #119–fix(server): announce each transaction of a batch as soon as it commits #126 are not involved. They are still 1.81× (S1) and 1.72× (MIX) v0.7.0.
• Tag filter matching 50 % of 100k messages: 34.6 ms p50, target ≤ 15 ms. A 1 % tag takes 1.1 ms.
• has_attachments=true takes 7.4 ms p50 with 0 hits on 100k messages, which looks like a full scan with no index behind it.
• Memory around the migration. The migrating process peaks at 538 MB RSS (100k small messages). Afterwards the process keeps a 32 MB (S1) or 95–152 MB (S2) footprint until it restarts, against about 5 MB after a normal restart.
• Peak RSS during ingest is 57–63 MB, against 16–23 MB on v0.7.0. It already shows at 4289944, so it predates phase 4.
• Delete-all on 100k messages takes 1.17 s (target 1.0 s); that is tracked in perf(storage): delete-all takes ~1.7 s on 100k messages #106.