Walter Harms pointed out the locking was rather obtuse in dupfd. This
patch makes it much clearer.  Present code acquires a lock in dupfd()
callee locate_fd(), and releases it in allocate_fd() OR dupfd() itself.

This changes cleans this up to acquire and release the lock entirely
within dupfd(), which moves locking completely out of locate_fd()
and allows the manual inlining of allocate_fd() into its only caller,
dupfd().

This change is functionally equivalent to the current code
[well, dupfd might run a cycle or two faster if the compiler optimizes well...]