fix: unlock l.mu before calling OnPotentialDeadlock callback

OnPotentialDeadlock is called while l.mu is held. If the callback
panics (e.g. to allow the goroutine to unwind while keeping other
goroutines running), l.mu remains locked, deadlocking all future
lock operations.

Unlock l.mu before calling the callback in both the recursive
lock and order violation detection paths. For the order violation
path, re-acquire l.mu afterward if the callback returns normally.

Fixes #29

Author: Yanhu007

deadlock.go

@@ -427,8 +427,9 @@ func (l *lockOrder) preLock(stack []uintptr, p interface{}) {
 						buf.Flush()
 					}
 					Opts.mu.Unlock()
+					l.mu.Unlock()
 					Opts.OnPotentialDeadlock()
-					break
+					return
 				}
 			}
 			continue
@@ -454,7 +455,9 @@ func (l *lockOrder) preLock(stack []uintptr, p interface{}) {
 					buf.Flush()
 				}
 				Opts.mu.Unlock()
+				l.mu.Unlock()
 				Opts.OnPotentialDeadlock()
+				l.mu.Lock()
 			}
 			// Copy both stacks: they're backed by pooled buffers that will be
 			// recycled in postUnlock, but l.order entries persist until MaxMapSize.