Don't grow lock extents too much if there are many conflicting lockers. For

Don't grow lock extents too much if there are many conflicting lockers.  For
write locks we grow to at most 32MB upwards and never downwards.  For read
locks we also limit downward growth under contention to try and avoid granting
overlapping locks (cancellation of which is very slow for large files).

This actually improves single-file IOR write performance noticably
(100-500MB/s) with no impact to FPP IOR, but we still suffer on the read
side as eventually the read locks overlap and cancellation is very CPU
intensive.  Fixing bug 20 is the solution to that.
b=2919
r=phil