fix(vmm): persist guest events off the executor and outside the global lock

[kvinwang]

Problem

HostApi.Notify persisted guest events on an async worker while holding the global VMM state mutex. An instance.info file/directory fsync could therefore stall unrelated VMs' control-plane operations.

Fix

• Keep in-memory event updates under short state locks; move persistence to spawn_blocking.
• Serialize writes with the existing per-VM lifecycle lock, owned by the blocking task even after request cancellation.
• Reject notifications targeting a removed/replaced VM. Ordinary progress events do not wait for lifecycle I/O.

Fixes #1269.

Validation

Linux x86_64, Rust 1.92.0, debug tests; no QEMU/TDX VM or storage-throughput benchmark:

• cargo test -p dstack-vmm: 212 passed, 3 existing root/network-namespace tests ignored.
• cargo fmt -p dstack-vmm -- --check and CI-style production clippy passed.
• A single-thread Tokio test stalls persistence on a channel (5-second failure guard), confirms other events/state access stay responsive, cancels the request, and checks lifecycle exclusion plus ordered subsequent writes.
• Temporary-directory tests check persisted contents, shutdown markers, error propagation, and stale-VM rejection.

This removes executor/global-lock occupancy by persistence; it does not make fsync itself faster or claim a production speedup percentage.