[TRTLLM-12417][feat] Exact multimodal cache hashing + EPD unlock for videos

[venkywonka]

Description

This PR makes multimodal prompt-layout bookkeeping exact for non-contiguous multimodal token layouts.
The PR also pipecleans metadata transport to enable video support for EPD disagg.

The legacy cache contract reduced each multimodal item
to a single approximate prompt span. That loses the real prompt geometry when an
item is represented by multiple prompt-token runs or when special wrapper tokens
belong to the multimodal layout but do not consume encoder-output rows.

The PR introduces multimodal_item_runs as the canonical internal layout. Each
multimodal item carries ordered contiguous prompt-token runs, plus local
non_embed_offsets for multimodal wrapper tokens.

Runtime code derives
embedding lengths, prompt lengths, cache-key spans, disaggregated prompt
expansion, and backend request metadata from that canonical layout.

Goals

• Represent multimodal prompt layout as per-item contiguous runs instead of a
  lossy single start/length span.
• Preserve item-run metadata through Python preprocessing, C++ request types,
  nanobind bindings, executor request construction, serialization, Triton
  backend request transport, and disaggregated handoff.
• Generate multimodal KV-cache extra keys from exact item runs so text or
  framing gaps do not inherit multimodal content hashes, while later sparse
  multimodal islands still get the correct hash and item offset.
• Expand Qwen2-VL and Qwen3-VL disaggregated prompt token IDs from item runs for
  video layouts where markers do not one-to-one map to contiguous embeddings. This unlocks EPD for videos.
• Keep OpenAI JSON disaggregated parameters scoped to JSON-safe state for this
  PR; multimodal disaggregated JSON transport remains a follow-up.

Review Map

• Core item-run contract and producers:
  tensorrt_llm/inputs/, multimodal modeling helpers, and focused input tests.
• Runtime ABI and serialization:
  C++ executor/batch-manager request types, nanobind bindings, pyexecutor
  request plumbing, MultimodalInput serialization, and serialization tests.
• Cache correctness:
  C++ block-key generation, KV-cache manager tests, Python V2 resource-manager
  cache-key estimation, cache-event tests, and item-run offset handling.
• Disaggregated EPD propagation:
  DisaggregatedParams, LLM preprocessing, result handoff, Qwen VL prompt
  expansion, and disaggregated item-run tests.
• Backend transport:
  Triton inflight batcher utilities and tests for validated item-run transport.
• AutoDeploy boundary:
  Qwen3.5 MoE VLM custom model/shim paths and tests for single-run item staging
  plus explicit sparse same-item rejection.
• Scope cleanup:
  OpenAI protocol/server paths now avoid exposing non-JSON-safe multimodal
  disaggregated state in this PR.

Test Coverage

• C++ focused coverage:
  • blockKeyTest
  • kvCacheManagerTest
  • multimodalInputTest
  • serializeUtilsTest
  • Triton inflight batcher utilsTest
• Python focused coverage:
  • multimodal input item-run validation and derived metadata
  • multimodal embedding length derivation
  • runtime and executor binding coverage
  • Python KV-cache estimation and V2 item-run cache keys
  • disaggregated item-run propagation and Qwen VL prompt expansion
  • LLM KV-cache event multimodal input coverage
  • AutoDeploy Qwen3.5 model/shim coverage
• Validation already performed on this branch family before the final history
  rewrite:
  • Python focused gates: 101 passed, 158 deselected
  • Native focused gates: 9 passed
  • git diff --check passed
  • Qwen2-VL-2B MMMU smoke matched baseline: feature accuracy 40.0, baseline
    accuracy 40.0, mm_keys=654
  • Qwen3-VL EPD C/G service smoke on Lyris passed for both baseline and
    feature; both returned the same multimodal response

PR Checklist

Please review the following before submitting your PR:

• PR description clearly explains what and why.

• PR follows TRT-LLM CODING GUIDELINES to the best of your knowledge.

• Test cases are provided for new code paths.

• Any new dependencies have been scanned for license and vulnerabilities.

• CODEOWNERS updated if ownership changes.

• Documentation updated as needed.

• Update tava architecture diagram if there is a significant design change in PR.

• The reviewers assigned automatically/manually are appropriate for the PR.

• Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.

[venkywonka]

closign this in favor of #13815 and #13864