Fix timings of completion, add timing of sampling back

[lucasavila00]

Now that we're sampling fully in CPU, we should not merge the sampling timings into completion timings.

This will likely show an improvement on mistralrs-bench's tg test.

Notice llama-bench selects a random token https://github.com/ggerganov/llama.cpp/blob/master/examples/llama-bench/llama-bench.cpp#L1161 so currently tg test is not a truly fair comparison.

Basically, we need to revert https://github.com/EricLBuehler/mistral.rs/pull/151

[lucasavila00]

I'm not sure if it's better to re-design the code or mutate some variable here, which is the point where we know the correct timings

[EricLBuehler]

I think that we should start sampling timing there as it excludes the sync point.

[EricLBuehler]

Done in a previous PR which didn't close this.