Store tokens in u32 instead of usize

[tmm1]

based on upstream commit openai@c2960c1

cc openai#251

[tmm1]

Running benches/bench.rs (target/release/deps/bench-c2328a0362693adb)
cl100k_base_estimation  time:   [1.9916 ms 1.9949 ms 1.9987 ms]
                        change: [-0.3951% -0.1311% +0.1318%] (p = 0.33 > 0.05)
                        No change in performance detected.
Found 16 outliers among 150 measurements (10.67%)
  7 (4.67%) high mild
  9 (6.00%) high severe

cl100k_base             time:   [14.279 ms 14.297 ms 14.317 ms]
                        change: [-3.6656% -3.5011% -3.3338%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 6 outliers among 150 measurements (4.00%)
  2 (1.33%) high mild
  4 (2.67%) high severe

cl100k_base_ordinary    time:   [14.451 ms 14.476 ms 14.506 ms]
                        change: [-1.0329% -0.8089% -0.5677%] (p = 0.00 < 0.05)
                        Change within noise threshold.
Found 10 outliers among 150 measurements (6.67%)
  2 (1.33%) high mild
  8 (5.33%) high severe

cl100k_base_50atatime   time:   [17.989 ms 18.002 ms 18.017 ms]
                        change: [-3.0768% -2.9049% -2.7420%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 4 outliers among 150 measurements (2.67%)
  2 (1.33%) high mild
  2 (1.33%) high severe