server : add last-5-seconds generation speed display

[akx]

Overview

This PR adds a tg(last 5s) display to the server token generation speed log.

I noticed tg is computed as an average over the entire computation, which, for instance on my Macbook, doesn't really tell the full story (other than that yes, speed decreases as the context grows longer).

Example:

6.33.268.728  I slot print_timing: id  0 | task 0 | n_decoded =    100, tg =  12.24 t/s, tg(last 5s) =   0.00 t/s
[...]
7.55.294.463  I slot print_timing: id  0 | task 0 | n_decoded =   1005, tg =  11.14 t/s, tg(last 5s) =  10.77 t/s
[...]
11.19.803.987 I slot print_timing: id  0 | task 0 | n_decoded =   2888, tg =   9.80 t/s, tg(last 5s) =   6.98 t/s
[...]
12.49.318.567 I slot print_timing: id  0 | task 0 | n_decoded =   3415, tg =   8.89 t/s, tg(last 5s) =   5.49 t/s
12.52.326.484 I slot print_timing: id  0 | task 0 | n_decoded =   3430, tg =   8.86 t/s, tg(last 5s) =   5.22 t/s
12.55.413.214 I slot print_timing: id  0 | task 0 | n_decoded =   3446, tg =   8.83 t/s, tg(last 5s) =   5.10 t/s
12.58.511.666 I slot print_timing: id  0 | task 0 | n_decoded =   3463, tg =   8.80 t/s, tg(last 5s) =   5.47 t/s
[...]

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES: assisted by Claude Code (Opus 4.8).

[ngxson]

IMO a sliding window is quite overkill. a rolling windows that is cleared each 5s will be much more simple

[akx]

@ngxson Sure! Replaced with a static 5-second window, no extra allocations.

[ngxson]

cc @ggerganov : not exactly sure if you think this feature would be useful ?

[ggerganov]

I think it's useful to have this info.