This eliminates some overhead, because the compiler doesn't need to
consider these functions throwing.

nfs_destroy_context() will invoke all pending callbacks with
err==-EINTR.  In CancellableCallback::Callback(), this will invoke
NfsConnection::DeferClose(), which however is only designed to be
called from nfs_service().  In non-debug mode, this will leak memory
because nfs_close_async() is never called.

Workaround: before nfs_destroy_context(), invoke nfs_close_async() on
all pending file handles.

If an async opertion is in progress, nfs_close_async() will make
libnfs crash because the RPC callback will dereference an object that
was freed by nfs_close_async().

Close the newly allocated file handle passed to the callback.

All locks are currenly held from only a single thread (the IOThread)
and thus we don't need the Mutex.

Not used anymore because NfsManager now uses boost::intrusive::set
instead of std::map.

This allows eliminating the indirection code from the NfsConnection
destructor.

Reduce Remove() overhead because we don't have to walk the list to
find an iterator by reference.

std::map::emplace() is only available from gcc 4.8 on.

More robust and cancellable.