Add a network-contention benchmark to the demo harness

[Fiona2016]

Why

0.5.0 shipped three pacing settings and advice on when to reach for them, but
nothing measured whether upload traffic actually delays an app's own requests, or
by how much. The advice rested on the mechanism being plausible.

It holds up. Against a silent-SDK baseline at 2.7 Mbps, the default pacing costs
+59% at p95, and RARE + LARGE + LOW costs +7%.

Group (2.7 Mbps, median of 3 runs)	mean	p95	vs baseline
SDK silent	58 ms	59 ms	—
default (AVERAGE/MEDIUM/MEDIUM)	66 ms	94 ms	mean +13.8% · p95 +59.3%
low (RARE/LARGE/LOW)	57 ms	63 ms	mean −1.7% · p95 +6.8%

Cross-checked at 0.5 Mbps: +53.1% against +20.2% at p95. Same direction, and the
effect sits in the tail rather than the mean — the shape reported for the Android
investigation.

What

A netbench scenario on the existing e2e harness rather than a separate project:
the mock intake already serves both the SDK's uploads and a stand-in origin for
the app's requests, which is exactly the shared link the experiment needs. Runs
are driven by aa start --ps and collected from HiLog, so a full run is
scriptable over hdc — emulator or real device, no tapping.

UPLINK_BPS=2700000 node scripts/e2e/mock-intake.mjs 19533 /tmp/capture.ndjson
hdc shell aa start -a EntryAbility -b com.flashcat.sdk.harmony \
  --ps e2e_endpoint http://10.0.2.2:19533 --ps e2e_scenario netbench \
  --ps netbench_pacing low --ps netbench_rounds 40
hdc shell "hilog -x" | grep NETBENCH

Throttling is opt-in via UPLINK_BPS, so existing smoke runs are untouched.

Notes for review

The server models a link, not a quota. A token bucket sized to a second of
bytes was the first attempt and it measured nothing: the bucket starts full, so a
request smaller than the bucket draws its bytes instantly and never queues. Real
links have no burst credit — they serialize at the line rate, and a request
arriving mid-transmission waits. That queueing is the effect under test.

Groups are whole runs, not alternating arms. Alternating uploads on/off
inside one run reads as the stronger design, but consent PENDING stops new drains
rather than an upload already in flight. Once an upload outlives its round it
spills into the supposedly quiet arm and both arms carry traffic. At 0.5 Mbps
that erased the effect entirely — not because contention had gone, but because
the baseline was contaminated. Drift is handled by repetition and medians.

No flush(). Forcing an upload bypasses the batch window, the cycle interval
and the per-cycle bound — every setting the run exists to compare. Load is
organic and the SDK's own scheduler decides when to upload.

Verification

• ./scripts/ci-check.sh — ALL CHECKS PASSED (148 tests, HAR build, demo HAP
  build, ArkTS unit-test type check).
• Results above reproduced after the final refactor: off 56/61, default 66/100,
  low 58/66.
• Measured on the DevEco emulator. The emulator reproduces contention and the
  relative ordering faithfully, but its queue is cleaner than a real radio, so
  the absolute figures should not be quoted externally without a device run.