Make AdapterManager::add_peripheral idempotent

[DaveMeade]

add_peripheral asserts the peripheral is absent and then inserts it. That is a
check-then-act pair on a concurrent DashMap, and the caller that reaches it is
inherently racy: droidplug's Adapter::report_scan_result looks a peripheral up,
takes the None branch, and only then calls add, holding nothing across the two
steps.

Two scan results for the same device can therefore both observe None and both
add, and the second aborts the process. It is easy to hit on Android because
Manager::adapters() returns a process-global singleton adapter: an application
that starts a second scan, directly or through a library that scans internally to
find a peripheral, re-reports devices already in the shared map.

The abort is unusually hard to diagnose. These callers are spawned tasks, so Tokio
stores the panic payload in a JoinHandle nobody joins, and Android's default
panic hook writes to stderr, which is not in logcat. The observable symptom is a
task that silently stops existing, and every layer above reports a plain timeout.

The map keeps the entry it already has rather than replacing it: that one may
carry connection state a freshly constructed wrapper for the same address would
not. droidplug::Adapter::add returns the map's instance for the same reason, so
a wrapper that loses the race cannot shadow the live one.

Found against 0.12.0 in an app that scans on Android, and carried on a fork since
August. cargo check and cargo clippy --target aarch64-linux-android are clean
on master.

Two things I would take direction on:

• whether you would rather fix the race at the report_scan_result call site than
  make add_peripheral idempotent;
• whether clear_peripherals and DeviceDisconnected interact with a concurrent
  scan in ways this does not cover.

Happy to add a CHANGELOG entry if you want one.

🤖 Generated with Claude Code

[qdot]

Ok, this came into dev via a rebase, but I had to add a follow-on commit because your claude went wild on the comments, and also because this fixed the existing android issue while adding a new race to windows.

Also, we're on 0.13.x now, you'll want to update to that for future use.