Investigate the use of a value pool to reduce object creation

[halostatue]

Based on work by Charlie Somerville (@charliesome); found at SamSaffron@ca0bce0 and SamSaffron@6b34ec6.

[halostatue]

I’ve brought the changes in on a branch that will be pushed after 3.0 is released—initial benchmarking looks reasonable, but there’s a lot more allocations in exchange for a lot fewer objects (about 8k fewer objects in both full and columnar cases, which is fairly significant).

    Before
              Full      Columnar    Columnar Full
Allocations   212,916   124,021     178,153
Objects       54,792     30,307      30,307

    After
              Full      Columnar    Columnar Full
Allocations   234,602   142,975     197,113
Objects       46,277     21,880      21,880

                                     user     system      total        real
Normal                           6.770000   0.150000   6.920000 (  6.967850)
Columnar                        10.190000   0.090000  10.280000 ( 10.347297)
Columnar Full                   11.430000   0.120000  11.550000 ( 11.589130)
Cache Initialize                 0.220000   0.000000   0.220000 (  0.239905)
Cached                           4.320000   0.070000   4.390000 (  4.646856)
Columnar Cache Initialize        0.260000   0.000000   0.260000 (  0.265293)
Columnar Cached                  2.270000   0.040000   2.310000 (  2.594322)

In the full (json) case, the performance over 50 loads (rake benchmark:load) is about a full second slower (~20ms for a single load), but it has no apparent impact in the columnar cases.