fix(ds4): cancel abandoned inference

[FiloSpaTeam]

Propagate gRPC cancellation into DS4 prompt synchronization and poll it at decode boundaries.

Stop on failed stream writes and skip parser finalization and KV persistence for abandoned partial requests.

Assisted-by: Codex:gpt-5.6-sol

Description

This PR fixes DS4 inference continuing after the originating gRPC request has been cancelled or disconnected.

Cancellation is propagated into prompt synchronization and checked at safe decode boundaries, between accepted MTP tokens, during distributed route waits, and after streaming write failures.

Abandoned requests do not finalize incomplete parser state or persist partial KV-cache state. Backend failures retain their existing error status rather than being incorrectly converted to cancellation.

Notes for Reviewers

• Prompt synchronization uses a scoped cancellation callback that is cleared afterward.
• Distributed route waits poll for cancellation instead of blocking for the complete timeout.
• Streaming finalization is transactional: parser state and KV cache are committed only after the final write succeeds.
• Already-streamed chunks cannot be retracted.
• An active GPU kernel cannot be interrupted mid-execution; cancellation takes effect at the next safe boundary.
• Generation-limit and MTP-budget changes are intentionally handled in a separate PR.
• Both PRs modify grpc-server.cpp, so this branch may require a small rebase after the generation-boundary PR merges.
• All 17 standalone DS4 C++ test files pass.
• A complete CUDA/GB10 Docker end-to-end test was not available in the development workspace.

Signed commits

• Yes, I signed my commits.
• Documentation updated (docs/content/) for user-facing changes, or not applicable