Optimize copy_cells, part 2

[joe-maley]

This patch optimizes the copy_cells path for multi-fragment reads. The
following benchmarks are for the multi-fragment read scenario discussed
offline:

// Current
Read time: 4.75082 secs
  * Time to copy result attribute values: 3.54192 secs
    > Time to read attribute tiles: 0.311707 secs
    > Time to unfilter attribute tiles: 0.370434 secs
    > Time to copy fixed-sized attribute values: 0.898421 secs
    > Time to copy var-sized attribute values: 0.954925 secs

// With this patch
Read time: 3.04627 secs
  * Time to copy result attribute values: 1.83972 secs
    > Time to read attribute tiles: 0.274928 secs
    > Time to unfilter attribute tiles: 0.38196 secs
    > Time to copy fixed-sized attribute values: 0.517415 secs
    > Time to copy var-sized attribute values: 0.461847 secs

For context, here are the benchmark results for the single-fragment read. The
stats are similar with and without this patch:

Read time: 1.86883 secs
  * Time to copy result attribute values: 1.19411 secs
    > Time to read attribute tiles: 0.304055 secs
    > Time to unfilter attribute tiles: 0.351332 secs
    > Time to copy fixed-sized attribute values: 0.289661 secs
    > Time to copy var-sized attribute values: 0.142405 secs

This patch does three things:

1. Converts the offset_offsets_per_cs and var_offsets_per_cs in the var-sized
   path from a 2D array (vector<vector<uint64_t>>) to a 1D array (vector<uint64_t).
   The big win is construction and destruction time for the nested vector
   elements.
2. Partitions cell copying for both fixed and var-sized paths. The motivation
   is to reduce contention on the TBB threads and minimize time spent context
   switching between the individual cell slab copies.
3. Added a context cache for the fixed size path, similar to the existing
   var-sized context cache.