fix(pytorch): invalidate builder-flash-attn stage cache

Adds an inline comment inside the builder-flash-attn RUN block to
change its BuildKit stage hash, forcing a fresh source build against
the current torch 2.13 pin.

Background: earlier CI cycles on this branch cached a flash-attn wheel
that was compiled against torch 2.12 (undefined symbol
c10::impl::cow::materialize_cow_storage at import time on torch 2.13).
The bad wheel persisted through S3 delete + ECR tag delete because
BuildKit's stage cache lives across untagged image layers and the
fleet host's local buildkitd cache, both of which we cannot enumerate
and clean cleanly. Changing the RUN command text is the only way to
guarantee BuildKit re-executes the stage.

The comment is intentionally verbose and self-documenting so future
readers understand its purpose and don't reflexively delete it.

Author: Bhanu Teja Goshikonda

docker/pytorch/Dockerfile.cuda

@@ -89,6 +89,12 @@ ARG MAX_JOBS
 # docker/pytorch/${FRAMEWORK_SHORT_VERSION}/wheels/ is created by CI (fetch_cached_wheels.sh); may not exist locally.
 COPY docker/pytorch/${FRAMEWORK_SHORT_VERSION}/cuda/pyproject.toml docker/pytorch/${FRAMEWORK_SHORT_VERSION}/wheel[s]/ /tmp/wheels/
 RUN --mount=type=cache,target=/root/.cache/uv \
+  # Cache invalidation nudge 2026-07-14: prior BuildKit stage cache on this \
+  # branch contained a flash-attn wheel compiled against torch 2.12 ABI \
+  # (undefined symbol c10::impl::cow::materialize_cow_storage on torch 2.13). \
+  # This comment changes the RUN text hash so BuildKit re-executes the stage \
+  # against the current torch 2.13 pin. Do NOT remove until poisoned cache \
+  # layers are GC'd out of ECR + fleet host BuildKit local caches (~30 days). \
   mkdir -p /tmp/built_wheels && \
   WHL=$(find /tmp/wheels -name "flash*attn*.whl" 2>/dev/null | head -1) && \
   if [ -n "${WHL}" ]; then \