feat(website): add /benchmarks page with netem WAN harness numbers

[pvg13]

Summary

Publishes the latest snapshot from the netem-driven WAN harness on the website at `/benchmarks`. New page wired into the top nav alongside Docs / Examples / Demo.

What's on the page

Methodology — how each profile is shaped (`tc netem` inside Docker containers) and what we measure: time-to-first-sync, p50/p95 steady-state propagation over 50 writes, recovery time after a 5s blackhole.

Profile matrix — 12 profiles from `ideal` to `satellite`, populated with the most recent run's numbers. Source profiles span LAN, WiFi (home / busy / distant), mobile (5G / 4G fair / 4G bad / 3G), satellite, and a lossy LAN baseline.

```

Profile	First sync	p50	p95	Partition recovery
ideal	16 ms	16 ms	22 ms	2266 ms
lan_fast	23 ms	69 ms	71 ms	505 ms
ethernet_gigabit	24 ms	69 ms	71 ms	1776 ms
wifi_home	104 ms	95 ms	98 ms	512 ms
wifi_busy	160 ms	172 ms	213 ms	3619 ms
wifi_distant	798 ms	616 ms	824 ms	2191 ms
mobile_5g	216 ms	167 ms	198 ms	992 ms
cellular_fair	544 ms	357 ms	519 ms	3529 ms
cellular_bad	1973 ms	1761 ms	2718 ms	4098 ms
mobile_3g	759 ms	769 ms	1497 ms	1075 ms
satellite	4607 ms	3813 ms	5046 ms	>240 s †
lossy_lan	26 ms	71 ms	1111 ms ‡	2695 ms
```

Footnotes:

• † satellite partition-recovery exceeded the 240 s test cap on this run; high-RTT profiles have high run-to-run variance. Earlier runs recorded ~4.5 s.
• ‡ lossy_lan p95 is 15× p50 — filed as Engine: 15× p95 tail inflation under 5% packet loss (lossy_lan) #38.

Extended scenarios — long-offline, container-restart, asymmetric peer pairs. The `long_offline_30s` on cellular_bad row reflects #39 (1 of 10 writes fails to converge under heavy loss + long offline).

DCUtR section — explains the engine's hole-punching support, links to #51 for why the local Docker bench doesn't yet exercise DCUtR end-to-end.

Run-it-yourself — exact command for reproducing the numbers locally.

Implementation

Numbers are hardcoded as `&[ProfileRow]` / `&[ScenarioRow]` const arrays in `website/src/pages/benchmarks.rs` so they're type-checked and trivially diffable. To refresh: re-run the matrix locally, update the arrays, push.

Test plan

• `cargo build -p wavesync-website` — native build passes.
• `cargo build -p wavesync-website --target wasm32-unknown-unknown` — wasm32 build passes (this is the build that ships).
• Page renders sensibly in the desktop server-side build (verified via cargo build of the dioxus-fullstack target).

Related

• Numbers from netem harness landed in test(e2e): add netem-driven WAN benchmarks; remove e2e from CI #37 / test(e2e): NAT-topology shapes via iptables (closes #42) #50 / test(e2e): NatProfile::PortRestrictedConeAutoNATOk variant #52
• Engine signals filed as Engine: 15× p95 tail inflation under 5% packet loss (lossy_lan) #38, Engine: write loss + slow recovery under long offline + bad cellular (long_offline_30s on cellular_bad) #39
• DCUtR pre-requisite for live numbers tracked as Test harness: add SymmetricNat shape (SNAT/DNAT-based) to actually exercise DCUtR #51