remove coldest when full

[bitfaster]

Fix cycle logic so that if all items are touched when a new item is added (thus the coldest is cycled back into warm) we simply discard the coldest cache item (last item in cold queue).

[bitfaster]

Hit rate tests

In summary, better in 2 scenarios, worse in 2 scenarios and equal in 1 scenario. Arc OLTP with a cache size of 250 is the only test with a significant degradation.

Wikibench (better)

CacheSize	Before	After
25	30.75	32.40881000373165
50	38.54	39.54920682561558
75	42.52	43.51437745987301
100	45.1	46.151614417006265
125	47.22	48.18985790615074
150	48.69	49.614582684121785
175	49.93	50.743290409091635
200	50.83	51.62156813535368

Glimpse (unchanged)

CacheSize	Before	After
250	12.83	12.834580216126351
500	26.13	26.134663341645886
750	39.43	39.43474646716542
1000	43.28	43.275145469659186
1250	49.93	49.925187032418954
1500	52.4	52.40232751454697
1750	53.5	53.49958437240233
2000	53.55	53.54945968412302

Arc OLTP (worse)

CacheSize	Before	After
250	15.6	11.649792975950204
500	27.12	23.37747293919455
750	32.29	30.91796159252635
1000	36.01	35.13096937575549
1250	38.68	38.00370838324336
1500	40.69	40.19668652128492
1750	42.48	41.81120063009698
2000	44.16	43.32496485787266

Arc Search (very slightly worse)

CacheSize	Before	After
100000	9.63	9.326894162265988
200000	18.70	18.400852233908196
300000	26.83	26.678927981505275
400000	34.39	34.381865297163486
500000	41.43	41.726245393779095
600000	48.26	48.69219345890119
700000	54.61	55.08680008937266
800000	61.3	60.86261196113873

Arc Database (better)

CacheSize	Before	After
1000000	11.9	12.06719261894623
2000000	23.24	23.552243326784346
3000000	36.19	37.093629538181446
4000000	39.60	39.630505256620765
5000000	45.15	45.32610117487987
6000000	50.37	50.626977763792084
7000000	56.08	56.47311945854041
8000000	61.0	61.15599094556252

[coveralls]

coverage: 96.342% (+0.1%) from 96.219% when pulling 8445b79 on users/alexpeck/unbounded2 into 5565e86 on main.

[bitfaster]

With cold/warm reversed in the cycle method, the throughput test (multi-threaded) results in some runaway growth.