Split test=Speed into SpeedBulk and SpeedSmall and report weighted average for Small key speed test

[darkk]

I hope, that also addresses the goal that @wangyi-fudan had in #113 while being way more "stable" in computational terms.

It gives results looking like that:

$ ./SMHasher --test=SpeedSmall --extra SipHash
--- Testing SipHash "SipHash 2-4 - SSSE3 optimized" GOOD
...
Small key speed test -   64-byte keys -   231.26 cycles/hash
Average                                   151.602 cycles/hash
Average DNS query (| 99.8% of query.log)  123.685 cycles/hash
Average DNS name (| 99.4% of top-1m.csv)  118.621 cycles/hash
Average DNS name (| 99.6% of 200MiB.zone) 114.610 cycles/hash
Average UMASH (| 70.0% of startup-1M.bz2) 139.403 cycles/hash

or

$ SMHASHER_SMALLKEY_MAX=256 ./SMHasher --test=SpeedSmall --extra SipHash
--- Testing SipHash "SipHash 2-4 - SSSE3 optimized" GOOD
...
Small key speed test -  255-byte keys -   688.32 cycles/hash
Average                                   374.459 cycles/hash
Average DNS query (|100.0% of query.log)  124.329 cycles/hash
Average DNS name (|100.0% of top-1m.csv)  120.263 cycles/hash
Average DNS name (|100.0% of 200MiB.zone) 115.890 cycles/hash
Average UMASH (| 99.9% of startup-1M.bz2) 192.407 cycles/hash

Possible improvements are:

• computed distributions instead of tabulated ones might reduce size of SMHasher binary a bit, but 8 KiB of doubles is not that much anyway
• fit a linear model to extrapolate timings for keys having length greater than N instead of running a test, but I'm not that confident that it's useful unless we have real-world dataset with key-lengths being somewhat large