Explain the tradeoffs between different model and data parallelization strategies when scaling to multi-node GPU clusters at Zyphra.
