latency.

This saves a normal and a BufferPool allocation, ie. two mutex locks.

This tests multicore behavior.  Do not stat with more threads than
CPUs, since kernel+FUSE can never keep up with the stat threads.

Speedup 15% (lenovo T60, 2 CPUs, 2 stat threads)

FileSystem into the struct already is sufficient.

This fixes potential race condition with rename happening in parallel
to something else.