PyTorch optimizes jagged NestedTensor compile guards for faster reductions

PyTorch landed a performance optimization for its NestedTensor compile path, specifically targeting jagged (variable-length) tensors. PR #184053, authored by jansel and approved by oulgen, modifies the compile guard logic to skip redundant outer size and stride tracking when dealing with jagged NestedTensors. This reduces the per-call overhead for cached compiled functions, especially benefiting operations like reductions on batches of sequences with different lengths (common in NLP).

The optimization introduces fast-path handling for common metadata queries, so when a compiled graph is reused, it no longer wastes time recalculating dimensions that remain constant. The PR also adds regression tests for the guard set and ensures exact torch-function metadata dispatch, preventing future regressions. This is particularly impactful for users who rely on PyTorch's torch.compile with jagged NestedTensors, as it reduces latency in repeated forward passes. The change is part of ongoing efforts to make dynamic tensor shapes more efficient in PyTorch 2.x.