Fix Cuttlefish multi-instance restart deadlocks, socket hangs, and UWB/bootconfig configuration mismatches

[SuperStrongDinosaur]

Description
This change addresses a set of critical deadlocks, resource-leak hangs, and configuration-mapping bugs encountered in multi-instance CVD deployments, specifically during a cvd restart.

By resolving these architectural synchronization issues, single instances within a multi-instance group can now be safely restarted independently without hanging or crashing other running instances, and without deadlocking host-side shared daemons.

Fixes

1. Prevent FIFO Unlinking Deadlocks (SharedFD::Fifo & DeleteFifos )

Issue: Previously, SharedFD::Fifo always deleted the path before calling mkfifo(). In a multi-instance setup, global host-side daemons are started once and hold open connections to the VM instances. Unlinking these paths on a single-instance restart destroys the inode mapping, meaning the restarted instance's crosvm would construct new FIFOs that the active netsimd daemon has no knowledge of. This caused crosvm to hang indefinitely on startup, waiting for a reader/writer connection that would never come.

Solution: Modified SharedFD::Fifo to perform a stat() check on the target path first. If the file already exists and is verified to be a FIFO, it is opened directly instead of unlinking and recreating it. Removed bt_fifo_vm, nfc_fifo_vm, and uwb_fifo_vm from the unlinking sequence in ServerLoopImpl::DeleteFifos() to ensure their persistent paths are preserved across instance-specific lifecycles.

3. Ensure Proper FIFO Provisioning for Shared UWB Services (Files affected: uwb_connector.cpp)

Issue: The UWB host-connector logic only created the uwb_fifo_vm FIFOs if instance.enable_host_uwb_connector() was evaluated to true. If config.enable_host_uwb() was true but the specific instance did not run the connector, the missing FIFOs would crash host-side daemons or break crosvm initialization.

Solution: Decoupled the FIFO generation from the launcher command check. The FIFOs are now always initialized if global UWB is enabled, while the local host-connector service itself is only spawned if the instance enable_host_uwb_connector() is enabled.

4. Fix ProcessMonitor Shutdown Socket Hangs (Files affected: process_monitor.cc, process_monitor.h)

Issue: During a shutdown or restart event, ProcessMonitor could hang waiting on socket reads in ReadMonitorSocketLoop because the control channel remained blocked. If the socket connection returned an error or was half-closed, the loop could fail or crash rather than exiting cleanly.

Solution: Retained the child socket's raw file descriptor (child_sock_) within the ProcessMonitor class. Added an explicit child_sock_->Shutdown(SHUT_RDWR) call at the end of the MonitorRoutine execution to actively force-unblock any outstanding or stuck reads on the socket during shutdown. Hardened ReadMonitorSocketLoop to ignore read errors gracefully if the monitor has already been marked as shutting down (!running.load()).

Testing and Verification
The stability of these fixes was successfully validated under multi-instance conditions:

Multi-Instance Booting

Booted two parallel VM instances using the locally compiled binaries:
cvd create --config=sdv_core_instance1 --num_instances=2

Both instances initialized, completed guest execution, and reached the VIRTUAL_DEVICE_BOOT_COMPLETED signal cleanly.

Single-Instance Restarting (No Deadlocks)

Executed a single-instance restart command targeting instance 2:
cvd --group_name=cvd_1 --instance_name=2 restart

Instance 2 successfully halted its virtual processes, safely terminated crosvm and its children, preserved the shared Bluetooth/UWB/NFC FIFOs, and booted back to VIRTUAL_DEVICE_BOOT_COMPLETED.

Instance 1 remained completely undisturbed and functional during the entire restart cycle.

Verified using cvd status --print that both instances returned to a healthy Running state.

b/510634395