fix(pd): correct DeepSeek V4 layerwise cache handoff

[lucifer1004]

Motivation

DeepSeek V4 stores SWA, compressed KV, compressor state, and indexer state in independently paged cache groups. PD previously applied the legacy KV page indices to every contiguous buffer, so a Mooncake transfer could succeed while copying data between the wrong physical pages.

The DeepSeek V4 attention path also bypassed the common layerwise cache-readiness publication used by PD transfers.

Changes

• publish cache readiness from the DeepSeek V4 attention path
• associate each DeepSeek V4 contiguous buffer with its cache group
• carry source and destination group page tables and logical base offsets through the Mooncake control path
• match source and destination pages by logical position before constructing RDMA ranges
• preserve the existing flat-index path for ungrouped cache pools
• mirror group-aware address construction in the async manager without changing backend selection or wiring

Compatibility

Ungrouped cache pools retain their existing transfer behavior. Grouped-cache transfers require both peers to run the updated code; missing grouped metadata fails explicitly instead of copying incorrect pages.

This PR does not restore or enable mooncake_async executor wiring. Current executors continue to select the synchronous Mooncake managers.

Validation

• prek -C tokenspeed run --all-files
• 172 relevant runtime tests passed, with 1 skipped and 2 known SM120 CUDA cases deselected
• sync Mooncake layerwise PD qualified with TP2/EP2 and one prefill plus one decode replica
• full GSM8K: 93.63% flexible exact match and 93.48% strict exact match over 1,319 samples
• random 8K input / 1K output at concurrency 1 through 32: 252/252 requests completed
• no transfer, memory-registration, bad-address, CUDA, or OOM errors observed

The model qualification used the SM120 enablement stack from #648 plus this PD patch; those kernel changes are intentionally excluded from this PR.