What are practical strategies for reducing cold start latency in serverless Python functions?