fix(ds4): enforce generation boundaries

[FiloSpaTeam]

Clamp requested generation to the usable context after prompt sync while preserving the legacy 256-token fallback for omitted limits.

Constrain each speculative MTP cycle to the remaining request budget so accepted tokens cannot advance beyond the visible output limit.

Assisted-by: Codex:gpt-5.6-sol

Description

This PR fixes DS4 generation boundary handling in both streaming and non-streaming requests.

Explicit generation limits are clamped after prompt synchronization using the remaining context capacity. Omitted or non-positive limits retain the legacy 256-token fallback.

Speculative MTP decoding is also constrained by the remaining request budget, preventing accepted tokens from advancing beyond the visible output limit.

Notes for Reviewers

• The existing 256-token fallback remains unchanged.
• An explicit max_tokens: 32768 is treated as an upper bound and may be reduced to fit the available model context.
• Streaming and non-streaming paths use the same limit calculation.
• Request cancellation is intentionally excluded and will be submitted separately.
• All 17 standalone DS4 C++ test files pass.
• A full CUDA/GB10 Docker end-to-end build 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