Exact BPE token counts for OpenAI-family encodings (o200k_base for GPT-4o/4.1/5, cl100k_base for GPT-4/3.5/embeddings), single text or batches, with optional truncation to a token budget. Up to 512 KB per call.