trunk/4feff2113810b2f270676a3bd8f80d27bd31d5d6: [inductor] support is_inference flag for custom post grad pass (#171049)

The PyTorch development team has merged a significant technical commit to its main branch, introducing an 'is_inference' flag for custom post-gradient passes within its Inductor compiler. The change, identified as commit 4feff21 and resolving GitHub issue #170866, was approved by maintainer @karthickai. This update specifically modifies the Inductor's backend to allow user-defined optimization passes to detect the operational mode of a model. The core technical detail is that custom passes can now query a boolean `is_inference` flag, enabling them to apply different transformations or optimizations during the model's training phase versus its inference (deployment) phase. This is crucial because the optimal graph structure and memory layout often differ dramatically between these two contexts. For example, a pass might choose to aggressively fuse operators for latency during inference while preserving more granular operations for gradient computation during training. The context here is PyTorch's ongoing compiler evolution, with Inductor serving as its next-generation just-in-time (JIT) compiler for accelerating both eager-mode and compiled execution graphs. This change empowers framework developers and researchers building custom compilers on top of PyTorch, such as those working on model quantization, kernel fusion, or memory planning. The practical implication is more efficient AI models in production, as developers can now write a single pass that automatically tailors its behavior, potentially reducing code duplication and enabling more aggressive inference-specific optimizations that were previously unsafe to apply during training. This aligns with the industry-wide push to separate and highly optimize the inference pathway for cost and latency-sensitive deployments.