Reduce number of memcmp invocations in PTreeImpl::[upper|lower]_bound

[sears]

The storage server spends a large fraction of its time in PTreeImpl::lower_bound and PTreeImpl::upper_bound. These methods work by invoking T::operator<() in a few redundant ways.

Both methods populate a vector of tree nodes, invoking x < node->data each time they push a value on the end. At return, they invoke x < node->data again as they pop any nodes that are not less than / greater than x off the end of the vector.

lower_bound further computes node->data < x each time it pushes something onto the vector.

Reimplement the two methods to invoke the equivalent of memcmp(node->data, x) at most once for each node encountered during the traversal.

Note that, in the common case, operator< ends up compiling down to __memcmp_sse4_1. Be careful not to break that optimization.