getMaxConcurrency

Maximum number of concurrent VLM calls. Sizes the in-process context pool: the model weights are shared, each extra context costs only its KV/compute buffers and runs one more call in parallel (callers beyond the pool block for a free context).