Improve batched replace logic

[mlucy]

I was doing some profiling on an insert workload, and our batched replace code could use some work.

On a batched insert where the actual call to access->get_namespace_if()->write(...) takes about 50ms, you can expect to spend another 5-10ms just constructing/destructing the vector of datum_func_pair_ts. (The rest of the work, including JSON parsing, comes out to around 6ms.)

The situation is worse for actual batched replaces. As far as I can tell from looking at the code, if we have a batched replace containing 10 point replaces, we serialize the replacement function 10 times.

The batched_replace code needs to be reworked so that instead of always having a vector of <key,function> pairs, it has either a vector of <key,datum> pairs or a vector of keys with a single function.

[mlucy]

This is in next.