Reduce by key algorithm

[jszuppe]

This addresses issue #116.

This PR adds reduce by key algorithm (both serial and parallel version). Benchmarks for Boost.Compute, Thrust and BOLT are also included.

Performance test:

=== reduce_by_key with compute ===
size,time (ms)
2,0.154890
4,0.155084
8,0.155598
16,0.160392
32,0.171135
64,0.193295
128,0.232472
256,0.222300
512,0.727985
1024,0.722171
2048,0.712100
4096,0.709299
8192,0.711192
16384,0.713282
32768,0.744579
65536,0.813159
131072,1.015370
262144,1.135500
524288,1.864210
1048576,2.444740
2097152,3.390430
4194304,6.6587400
8388608,12.988500
16777216,31.645700
33554432,90.473600
=== reduce_by_key with bolt ===
size,time (ms)
2,0.462535
4,0.465382
8,0.466054
16,0.465169
32,0.465071
64,0.460544
128,0.460158
256,0.469034
512,0.487111
1024,0.464353
2048,0.435261
4096,0.471321
8192,0.469523
16384,0.480998
32768,0.501238
65536,0.556485
131072,0.631357
262144,0.732189
524288,2.281240
1048576,2.793540
2097152,2.672990
4194304,6.687320
8388608,12.697700
16777216,24.959100
33554432,53.312000

I checked in CodeXL and lower performance compared to the Bolt for big buffers is caused mainly by inclusive scan operation (used in reduce by key). At least on my HD7770.

[kylelutz]

Merged! Great work on this!

[jszuppe]

Thanks!