A token sampler for one row: (logits [1,V], step) → token array [1] on
device. Greedy is (l) => ops.argmaxAxis(l, -1); richer closures fold in
temperature / top-p / logits processors + history (built per request from
its sampler options, exactly like generate()'s sampleStep). Called once per
step 0..maxTokens-1 for this row (never beyond maxTokens).
A token sampler for one row: (logits [1,V], step) → token array [1] on device. Greedy is
(l) => ops.argmaxAxis(l, -1); richer closures fold in temperature / top-p / logits processors + history (built per request from its sampler options, exactly like generate()'s sampleStep). Called once per step 0..maxTokens-1 for this row (never beyond maxTokens).