fix(macos): drain autoreleased CoreWLAN objects in get_wifi_transmit_rate

[caiogondim]

Problem

On macOS, get_wifi_transmit_rate (src/os/macos/wifi.rs, called from
interface() to populate transmit_speed) queries CoreWLAN —
CWWiFiClient::sharedWiFiClient() → interfaceWithName → transmitRate —
without an enclosing autorelease pool. Those calls produce autoreleased
Obj-C / XPC objects.

When netdev is driven from a long-lived thread that has no run loop and no
draining autorelease pool, those objects are added to a top-level pool that
never drains, so they accumulate without bound. This happens in practice via
netwatch (and thus iroh), whose
interface monitor calls get_interfaces() on every network change.

Evidence

heap <pid> on a ~7h process using netdev through netwatch — CoreWLAN /
CoreWiFi objects dominate the heap (~73 MB):

count	bytes	class
298,213	33.4 MB	CWFRequestParameters (CoreWiFi)
223,658	17.9 MB	__NSXPCInterfaceProxy_CWFXPCRequestProtocolCoreWLAN
74,555	3.6 MB	CWFInterface (CoreWiFi)

RSS climbed ~18 MB/h, linear, no plateau. Linux hosts running the same software
were flat (their path is sysfs, no CoreWLAN), confirming the source.

Fix

Wrap the body in objc2::rc::autoreleasepool so each call drains its
autoreleased objects immediately (adds objc2 to the macOS deps for the pool
API). Behavior-preserving — the returned rate is unchanged; only the
autoreleased temporaries are freed per call.

Validation

Same software, only the patch differs:

• Before: CWFRequestParameters → 298,213 over 7h (~18 MB/h RSS climb).
• After: CWFRequestParameters pinned at 3, RSS flat (held 30+ min on a
  multi-host fleet, plus a separate long local run).

Reproducing

A unit test doesn't reproduce it — a synchronous in-process loop frees the
objects promptly (measured 0 growth over 100k calls); the leak is specific to
the async, threaded, over-time CoreWLAN XPC traffic on a pool-less long-lived
thread. To see it by hand on macOS: run a long-lived process that drives
netdev::get_interfaces() from a background monitor (e.g. via netwatch /
iroh) and watch heap <pid> | grep CWFRequestParameters climb; with the patch
it stays at a small constant.

[shellrow]

Thanks for the great fix and detailed investigation.

The change looks good. The only CI failure is a rustfmt issue, which I'll take care of on my side before the next release.