Implement 'faster on x% of runs'

[sweeneyde]

Based on #22:

Example:

py -m pyperf compare_to ..\cpython\floatbench.json ..\multiply\floatbench.json --table --win_percentage
+----------------+----------------------------+-----------------------------------------------+
| Benchmark      | ..\cpython\floatbench.json | ..\multiply\floatbench.json                   |
+================+============================+===============================================+
| int: x*x       | 28.5 ns                    | 27.7 ns: 1.03x faster (faster on 88% of runs) |
+----------------+----------------------------+-----------------------------------------------+
| float: x*x     | 28.4 ns                    | 27.9 ns: 1.02x faster (faster on 80% of runs) |
+----------------+----------------------------+-----------------------------------------------+
| int: x*...*x   | 262 ms                     | 220 ms: 1.19x faster (faster on 100% of runs) |
+----------------+----------------------------+-----------------------------------------------+
| float: x*...*x | 260 ms                     | 253 ms: 1.03x faster (faster on 90% of runs)  |
+----------------+----------------------------+-----------------------------------------------+
| Geometric mean | (ref)                      | 1.06x faster                                  |
+----------------+----------------------------+-----------------------------------------------+

I've not worked with this code before, so let me know what needs fixing.

[vstinner]

IMO it adds too much information which makes the whole output harder to read to users who are not used to this feature. I close this PR since it's inactive for 4 years. If someone wants this feature, I would suggest to enable it with a command line option.