feat(server): support KV Flash with same-backend target layer split

[weicj]

Summary

This PR extends KV Flash bounded KV residency to same-backend target layer split, so split targets can use a pool-sized KV cache instead of allocating full-context KV on every shard. The same-backend path is wired for Qwen35, Gemma4, and Laguna.

Changes

• Add KV Flash pool allocation and pager lifecycle support to Qwen35, Gemma4, and Laguna same-backend target layer split.
• Thread KvFlashPager through same-backend layer-split prefill/decode paths, including Qwen35 DFlash verify.
• Build layer-split attention graphs in pool-slot mode when KV Flash is enabled, including set-rows KV writes and slot-space attention masks.
• Align Gemma4 split boundaries so KV-sharing layers stay on the same shard as their source layer.
• Guard Laguna KV Flash mask uploads so only tensors allocated for the current graph are written.

Notes

• Local same-backend runtime smoke passed on dual Pro VII HIP with Qwen3.6-27B Q3, --target-devices hip:0,hip:1 --target-layer-split 1,1, and DFLASH_KVFLASH=1024; the request returned a valid OpenAI-compatible response.
• Local same-backend runtime smoke passed on dual Pro VII HIP with Gemma4 E2B Q4, --target-devices hip:0,hip:1 --target-layer-split 1,1, and DFLASH_KVFLASH=512; the server auto-aligned shards to [0,14)+[14,35) and returned a valid OpenAI-compatible response.
• Local same-backend runtime smoke passed on dual Pro VII HIP with Laguna-XS.2 Q4, --target-devices hip:0,hip:1 --target-layer-split 1,1, and DFLASH_KVFLASH=512; the pool was raised to 768 tokens for the SWA tail requirement and the request returned a valid OpenAI-compatible response.

[cubic-dev-ai]

3 issues found across 20 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>