fix(ic-agent): replace async-watch with tokio::sync::watch on non-WASM (0.46.1 release)

[lwshang]

Summary

Fixes #698 — [bug] failed to observe change after notificaton. panic reported against ic-agent 0.45 and still present in 0.46.0.

Root cause

async-watch v0.3.1 has a race condition in its changed() implementation: after the internal event-listener future resolves, maybe_changed() can return None if the executor context switches before the version is read, causing an unconditional expect(...) to panic. This race only manifests on multi-threaded tokio runtimes.

Fix

• Replace async-watch with tokio::sync::watch on non-WASM targets. tokio::sync::watch is battle-tested and race-condition-free. async-watch is kept for WASM where the single-threaded environment avoids the race.
• Change continue → break when fetch_receiver_recv returns Err (sender dropped). The old continue caused HealthManagerActor::run to spin-loop forever, starving the cancellation token arm in select!.

Changes

• ic-agent/Cargo.toml: move async-watch to WASM-only deps; add "sync" to non-WASM tokio features
• type_aliases.rs: platform-specific SenderWatch/ReceiverWatch type aliases
• dynamic_route_provider.rs: platform-specific watch channel construction
• health_check.rs: platform-specific fetch_receiver_recv helper; continue → break on closed channel; two regression tests

Regression tests

Two tests added to health_check.rs (non-WASM only):

1. test_health_manager_no_panic_on_rapid_updates_and_shutdown — 50-iteration stress test with multi-threaded tokio, flooding the watch channel 200× per iteration while health-check actors run at 1 ms intervals. Captures JoinHandle so any spawned-task panic surfaces as a test failure.

2. test_health_manager_exits_when_fetch_sender_dropped — deterministic test: drops fetch_sender while keeping the cancellation token alive, expects the actor to exit within 2 s. With the old continue bug this test times out (confirmed by temporarily reverting the fix and running the test).