trunk/c4df6c686a3a367b8dd5fde46d9e75d4bb1da22d: Move autograd_backward out of FX custom metadata (#180251)

The PyTorch team has resolved a subtle but significant bug in their FX graph tracer that was causing forward operations, specifically FlexAttention nodes, to be incorrectly tagged as backward operations. The issue stemmed from how the tracer stored the `autograd_backward` metadata in a shared dictionary within the node's custom metadata. When nested tracing scopes executed, this shared dictionary would get mutated in place, causing the `autograd_backward` flag to "leak" from actual backward nodes to unrelated forward nodes. This pollution of node metadata could lead to incorrect graph optimizations during compilation.

The fix, authored with assistance from Claude AI, involved treating `autograd_backward` as internal tracing state rather than user-facing custom metadata. The team implemented a dedicated `current_meta["autograd_backward"]` flag during tracing and used a `setup_stacktrace_preservation_hook` to properly stamp this metadata on newly created nodes. This architectural change ensures clean separation between internal tracer state and user annotations like `{"ac_region_id": 0}`. With proper node tagging restored, the regional inductor partitioner can now correctly reorder backward nodes and safely disable the rematerialization (remat) pass for compiled training graphs, as the pass is only needed before final graph decisions are made.

The team added regression tests covering both non-strict tracing scenarios and FlexAttention-specific cases to prevent future occurrences of this metadata leakage bug. This fix is particularly important for users leveraging PyTorch's `torch.compile` feature for whole training steps, as it ensures that graph partitioning and optimization passes behave correctly based on accurate node metadata.