Clarify partitioned CUDA output and failure contracts
Document automatically allocated outputs as shared mutable per-ID views and require independent copies to preserve results. Extend the staging regression to retain an earlier output across fresh-fetch and IOBinding runs while preserving a deep copy. Narrow failure invalidation to post-start partitioned errors and cover an eager-origin Reshape failure followed by successful existing captured replay.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 29a73897-8aa5-41e1-ab09-b695c52d121d
