Add a reproducible benchmark harness and a CI regression gate

[EONRaider]

Target: 1.4.0 (Tier 1 — earn the benchmark)

There is no benchmark anywhere in the repo — no timeit, no
perf_counter, no pytest-benchmark, no performance job in CI. The
Tier 1 work (#82, #83, #84, #85) is worth roughly 2.4× end to end,
and nothing would stop it silently regressing.

This issue closes Tier 1 and is the gate on publishing any performance
claim.

What to do

1. A benchmark script that walks the existing 97-frame corpus and
   reports frames/sec — the corpus is already the right workload
   because it exercises DNS, DHCP and IPv6 extension headers rather
   than one synthetic TCP frame.
2. Optional comparison mode against dpkt and scapy, with both as
   dev-only extras so the zero-dependency guarantee is untouched.
3. A CI job that fails on a significant regression against a committed
   baseline.

Why it is worth doing properly

Every published benchmark in this space is unusable: self-published by
the winner, run on a Core 2 Duo, or comparing CPython 3.6 against
CPython 2.7. A neutral, reproducible, CI-run comparison does not
currently exist for Python packet parsers, and publishing one is a
contribution independent of where NETProtocols places in it.

Acceptance criteria

• uv run one command reproduces the numbers.
• A regression beyond an agreed threshold fails CI.
• Numbers land in the README only after Hoist protocol dispatch into module-level tables #82–Cache DNS section parsing instead of re-serializing per access #85 have merged.