Deduplication tree optimizations

Deduplication tree can now be configured to be indexed
by full memcmp()s instead of hashes, hence reducing
footprint and avoiding the cost of a hash in each operation.

Of course there is a tradeoff between the cost of the hash and
the lookup time with memcmp()s. Experimentation so far suggests
hashless version is better so it's set by default.

Besides, some profiling functions are added (commented out).

A new metric NR_EVICTIONS_OK is added.

Signed-off-by: Jacobo Giralt <jacobo.giralt@bsc.es>