fix(tps): correct off-by-one in decode token count for generation TPS

[paul90317]

Overview

Generation TPS was computed using an extra decode token in the token count, while the decode time measurement did not include this extra step. This caused an inflated TPS value due to mismatched token/time accounting.

This change fixes the off-by-one issue in decode token counting to ensure consistent alignment between decode tokens and decode duration.

Additional information

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: yes (used for commit message drafting and wording)

[paul90317]

A potential issue may be in the stop condition that uses slot.n_decoded:

// check the limits
if (slot.n_decoded > 0 && slot.has_next_token && !slot.has_budget(params_base)) {
    slot.stop = STOP_TYPE_LIMIT;
    slot.has_next_token = false;
}

This may cause the model to output one more token than the configured limit.

I will investigate a better solution.