[yechank-nvidia]

Stacked on #20897 (bab3171542), after #20391. Review the handoff-only changes.

Description

Let prefill reuse an encoder's completed prompt/layout instead of rebuilding it. Encoded items and optional finalized layout use one versioned payload, with model-selected, P-owned state for decode.

Existing Dynamo remains compatible. Keep the existing DisaggregatedParams constructor fields, attributes, processor hook, and default result output. New consumers explicitly opt in:

legacy = encoded.disaggregated_params
versioned = encoded.export_multimodal_handoff()

The new method is a beta API addition. Both formats feed the common execution Params; postprocessing preserves the payload and each consumer's selected output. Existing APIs are not deprecated in this PR. Announce a transition period before a later removal.

Coordinated consumer: ai-dynamo/dynamo#17185, which also retains its older-backend fallback.

Test Coverage

• Matching native build, H200 NVL: 1,085 passed, 7 skipped, 35 deselected, plus 2 subtests.
• Existing executor/postprocessing regressions: 32 passed.
• Compatibility matrix: unchanged Dynamo 144 passed with this backend; updated Dynamo 148 passed against both the old API and the new API.
• Real Qwen3-VL-2B: AGG = legacy E/P = versioned E/P for one- and two-image requests, 12 generated tokens each; DeepStack width 8,192 preserved.
• Pre-commit and production mypy roots passed. Full configured mypy 1.19.1 reports the same 12 sampler-test diagnostics on both split branches. API stability reference updated for the additive beta method; committed APIs remain unchanged.

AIPerf, peak-memory comparison, and GPU-isolated Dynamo E/P/D remain pending. Local NIXL P/D initialization did not complete on this node and is not counted as a pass. No new model/modality support or performance gain is claimed.

PR Checklist

• Reviewed the applicable PR checklist items.