Fix up the spinlock situation around rq->rq_list and the lists it lives on.

We avoid a deadlock in ptlrpc_free_committed() when it calls ptlrpc_free_req().
Likewise when ptlrpc_cleanup_client() calls ptlrpc_free_req().

A note for Mike in ll_recover() for when it is re-enabled:

   We should not hold a spinlock over such a lengthy operation.
   If necessary, drop spinlock, do operation, re-get spinlock, restart loop.
   If we need to avoid re-processing items, then delete them from the list
   as they are replayed and re-add at the tail of this list, so the next
   item to process will always be at the head of the list.