Preserve unlimited token sentinel in text generation CLI

[sylvesterkaczmarek]

Summary

Pass the generation CLI's documented unlimited-token sentinel to the selected backend unchanged.

--limit 0 is the default and is documented as disabling the token limit, but gpt_oss.generate currently converts 0 to None. The Torch and Triton generators implement unlimited generation with integer 0 and compare num_generated_tokens < max_tokens, so None can cause a TypeError. The vLLM generator also accepts 0 and converts it internally for vLLM.

Fixes #268.

Fix

Pass args.limit directly as max_tokens.

Regression coverage

Adds a backend-stub test around main() verifying that the default/unlimited limit=0 reaches the token generator as integer 0 without requiring a checkpoint, GPU, or inference dependency.

Finite positive limits, stop tokens, temperature, logprob behavior, and backend selection are unchanged.