from report now it is not easy to know how much KVCache VRAM is saving

[gaowayne]

please review below data

================================================================================
Mode                      Context    Peak VRAM    KV Est     Speed        Prefill
================================================================================
fp16                      460        6098         159        68.9         0.294
turboquant-4bit           460        6066         170        32.9         0.321
turboquant-3bit           460        6070         173        40.1         0.038
fp16                      930        6304         303        73.3         0.019
turboquant-4bit           930        6242         343        33.1         0.039
turboquant-3bit           930        6249         350        39.7         0.034
fp16                      1860       6715         605        69.5         0.021
turboquant-4bit           1860       6590         687        32.6         0.054
turboquant-3bit           1860       6605         702        39.3         0.045
fp16                      3720       7532         1212       64.0         0.040
turboquant-4bit           3720       7285         1376       32.5         0.119
turboquant-3bit           3720       7317         1407       38.7         0.092