Reduce BPE merge update allocations and add file progress logs

[voidful]

Summary

• aggregate BPE pair count deltas and update positions per worker before updating global merge tables
• avoid collecting all per-word merge changes into a single intermediate vector
• add optional file preprocessing stderr progress controlled by TOKENIZERS_PROGRESS_LOG_INTERVAL_SECONDS

Testing

• cargo check
• cargo test models::bpe::trainer -- --nocapture

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.