Cherry-pick PR #368 async data path onto ww26_PR_async_data

[Copilot]

This ports the async non-GPU data path work from PR #368 onto ww26_PR_async_data, preserving commit order while skipping the empty planning commit and commits already present on the target branch. The branch diverged enough that the port also needed a thin compatibility layer and a small set of branch-local test updates.

• Async non-GPU store path

  • Ports the async store pipeline, including preemption-aware gather flushing, SHM direct-gather handling, background gather/commit execution, and the follow-up logging/profiling commits.
  • Carries over the dedicated AsyncDataTransferContext split and related worker/connector changes.

• Branch compatibility shims

  • Reconciles API/name drift on ww26_PR_async_data so the cherry-picked code can coexist with older branch-local interfaces.
  • Adds aliases for renamed transfer-context, protocol, key, and module symbols instead of rewriting the ported logic around branch-specific naming.

  IPCCacheEngineKey = IPCCacheServerKey
  RegisterNonGpuContextPayload = RegisterEngineDrivenContextPayload
  MPCacheEngineContext = MPCacheServerContext

• CPU / non-CUDA layout handling

  • Keeps the branch’s CPU-default HND safeguard, but allows explicit layout hints to override it where tests and non-CUDA transfer paths need deterministic NHD behavior.

• Multiprocess transfer test alignment

  • Updates the non-CUDA / SHM transfer tests to match the branch’s current module layout and runtime assumptions.
  • Renames and adapts server-side non-GPU transfer coverage to the ported interfaces, including chunk-subset gather behavior and SHM registration flows.

• Cherry-pick scope

  • Includes the PR Async data path #368 async-data commit series and skips commits that are already present on ww26_PR_async_data.
  • Keeps conflict resolution biased toward the PR Async data path #368 side, with only the minimal branch-local glue needed for the port to land cleanly.