Fix preprocessor reference logprob worker wiring

[taivu1998]

Summary

• assign real reference TrainableLLM clients to preprocessor workers at process startup
• add an early guard for KL-enabled preprocessing when no reference LLM URLs are available
• remove the dead per-chunk reference-selection path and reuse the existing tokenizer inside workers
• update launcher validation so final_kl_coef also requires reference-model capacity

Closes #62.

Problem

The preprocessor constructed reference LLM clients from me.llm_urls, but each worker process was still launched with llm=None. As a result, preprocess_dataset(...) always fell back to copying rollout logprobs into ref_logprobs.

That meant KL-enabled runs were no longer using the frozen reference policy that run_ref_llm(...) launches. Instead, the trainer compared the new policy against rollout-time actor logprobs, which changes both the loss semantics and the reported KL metrics.

What changed

• add small helpers in pipelinerl/preprocess.py to detect when reference logprobs are required, validate setup, and assign reference clients to workers
• pass one real reference TrainableLLM into each preprocessor worker when it is spawned instead of hard-coding None
• set llm.tokenizer = tokenizer inside the worker so the reference client can reuse the already-loaded tokenizer
• log the discovered reference URLs and each worker's reference source for easier runtime validation
• remove the unused next_llm_index code from the main preprocessing loop
• extend pipelinerl/launch.py config validation to treat either kl_coef or final_kl_coef as requiring world.preprocessor_fraction > 0

Validation

• python3.11 -m py_compile pipelinerl/preprocess.py pipelinerl/launch.py
• python3.11 -m compileall pipelinerl
• git diff --check

Notes

• only pipelinerl/preprocess.py and pipelinerl/launch.py are included in this PR
• analysis/plan markdown files in the working tree were intentionally left out of the commit
• I did not run a full distributed end-to-end training job here because that requires the full runtime stack and live reference/inference services

[taivu1998]

Hi @rizar, could you help review this PR? Thank you so much!