i2c: clear the controller on every IRQ, not just armed buses

A completed master write leaves bits set in I2CM14, and those bits hold
the controller's shared interrupt line asserted. handle_slave_interrupt()
is the only code that clears them, and it was reached only via
try_next_slave_event() behind an `if bus.notif_enabled` guard.

interrupt_ack() ran unconditionally, and it re-enables the NVIC line. So
on a bus that never arms slave notification the hardware source was never
cleared while the line was re-enabled, re-firing the interrupt forever.

Service the controller on every IRQ and gate only the latch and client
wake on notif_enabled. Buses that do arm notification are unaffected --
the call already ran for them.
diff --git a/services/i2c/server-runtime/src/lib.rs b/services/i2c/server-runtime/src/lib.rs
index 202123d..a51e6c5 100644
--- a/services/i2c/server-runtime/src/lib.rs
+++ b/services/i2c/server-runtime/src/lib.rs
@@ -121,8 +121,14 @@
             let irq = w.user_data as u32;
             let acked = w.pending_signals & irq_signals;
             if let Some(bus) = buses.iter_mut().find(|b| b.irq == irq) {
+                // Service the controller even with no client listening: a master
+                // write leaves I2CM14 latched, holding the shared IRQ line
+                // asserted, and only the driver clears it. Skipping this makes
+                // the `interrupt_ack` below re-enable an interrupt that is still
+                // asserted, so it re-fires forever.
+                let event = bus.driver.try_next_slave_event();
                 if bus.notif_enabled {
-                    match bus.driver.try_next_slave_event() {
+                    match event {
                         Ok(Some((kind, _))) => match kind {
                             I2cIsrEvent::SlaveWrRecvd => {
                                 // Master finished writing — drain buffer into latch.