fix(vllm): preserve hybrid block tables for external KV

[ketor]

Problem

vLLM's current hybrid memory allocator passes only the blocks allocated in the current scheduler step through NewRequestData.block_ids. The dfkv connector treated that field as the complete request block table. On GLM-5.3-Flash (MLA/DSA/KDA hybrid cache), this caused asynchronous saves to fail with short block tables on TP ranks.

Stateful groups add two related constraints:

• Their physical block tables are compact and retain only manager-selected state slots.
• Equal scheduler/group block sizes do not mean that every logical chunk has a physical state slot.

The old uniform-block fast path therefore generated keys that could never be stored, and lookup convergence discarded otherwise complete sparse stateful prefixes.

Fix

• Capture the complete per-group block table in update_state_after_alloc and use it for new-request metadata.
• Derive store/load masks from each vLLM cache manager over the full aligned prefix.
• Expand compact stateful block tables back to logical chunk positions before constructing GPUDirect descriptors.
• Treat a contiguous prefix as complete after every object selected by that same manager mask exists.
• Add regressions for scheduler-step-only block IDs, compact state tables, and equal-block-size stateful masks.

Validation

Runtime: vllm/vllm-openai:glm53-flash (g487ecf187), GLM-5.3-Flash native FP8, TP4, B200, dfkv v2.23.2 RDMA.

• Unit regressions executed inside the dedicated vLLM image: FINAL_TESTS_PASS.
• ruff check: all touched files passed.
• Before fix: only TP0 stored; TP1-3 failed with block table has 0 ids.
• After fix, no MTP: all 11 logical objects hit; 44 physical GETs across TP4, 40,728,192 bytes, zero failed keys.
• After fix, MTP5: 8 logical objects hit; 32 physical GETs across TP4, about 170 MB read, zero dfkv I/O errors.
• Cold cross-instance no-MTP request: 15,317 prompt tokens, 0.802 s end-to-end.

All tests used a source process followed by a separate cold target process.