Summary

Adds docs/how-to/tune-worker-pools.md, an operator how-to for the
data-plane performance-tuning environment variables introduced with the
off-task encrypt/decrypt worker pools and per-peer connected UDP (#91).
None of these knobs were documented outside source doc-comments. The
guide pairs with the existing tune-udp-buffers.md: the two cover
adjacent concerns (UDP socket buffers vs. per-stage worker queues), and
an operator tuning one will want to find the other. Also adds the guide
to the how-to index.

Closes #101.

What it documents

All of the startup-time runtime knobs, grouped:

• Worker pool sizing: FIPS_ENCRYPT_WORKERS, FIPS_DECRYPT_WORKERS
• Per-peer connected UDP: FIPS_CONNECTED_UDP, FIPS_MACOS_CONNECTED_UDP
• Send-path backpressure: FIPS_SEND_BACKPRESSURE_SLEEP_AFTER /
  _SLEEP_MICROS / _DROP_AFTER
• Perf profiler: FIPS_PERF, FIPS_PIPELINE_TRACE,
  FIPS_PERF_INTERVAL_SECS
• macOS send tuning: FIPS_MACOS_ORDERED_SENDER, _WORKER_STRIDE,
  _WORKER_BATCH, _SEND_FLOW_IDLE_MS, _SEND_PACE_MBPS,
  _SEND_PACE_BURST_BYTES, _NET_SERVICE_TYPE

Each knob lists its real default (read from source, including the
macOS/Linux split on the backpressure gates), what it does, and the
observability mode that would prompt tuning it. Includes the
massive-mesh FIPS_CONNECTED_UDP=0 escape-hatch note (one drain thread
per established peer) and a see-also linking the UDP receive buffer
ceiling to the decrypt worker queue capacity.

Review ask

@mmalmi this documents your #91 work, so I would value an accuracy pass
on the prose, especially:

• the hash-by-destination single-flow caveat (raising encrypt workers
  does nothing for a single flow)
• the yield, then sleep, then drop backpressure escalation, and that
  DROP_AFTER is the ceiling checked first
• the macOS batch history (the note that a batch of 32 could trigger TCP
  collapse on Wi-Fi)
• the QoS-marking regression note for FIPS_MACOS_NET_SERVICE_TYPE

One correction worth flagging: the doc uses 32768 as the decrypt worker
channel capacity (decrypt_worker.rs), not the 1024 that #101's
description listed. 1024 is the encrypt channel capacity; the two
differ.

Notes

Documentation only, no code changes. Defaults and behavior were read
from source at this branch's base.