refactor(server): share target layer-split runtime helpers

[weicj]

Summary

This PR refactors the target layer-split server path so new model adapters do not need to keep copying the same runtime shell. The existing Qwen35, Gemma4, and Laguna adapters all repeat shard/range/snapshot-backend setup and the autoregressive decode/sampling loop; this moves those shared pieces into server/src/common/layer_split_runtime.* while keeping model-specific loading, graph execution, cache layout, snapshot payloads, and EOS rules inside each adapter.

Changes

• Add server/src/common/layer_split_runtime.h/.cpp.
  • init_layer_split_runtime() centralizes target path validation, GGUF layer-count inspection, layer range calculation, shard metadata initialization, peer-access setup, and snapshot backend initialization.
  • run_layer_split_ar_decode() centralizes the AR decode loop, CPU sampling handoff, token emission, cancellation handling, and EOS termination.
• Update the Qwen35, Gemma4, and Laguna layer-split adapters to call the shared runtime helpers.
  • Qwen35 still owns DFlash draft loading, feature-ring handling, DFlash speculative decode, and its Qwen-specific forward path.
  • Gemma4 and Laguna still own their partial loaders, per-layer graph builders, cache/snapshot formats, and model-specific decode-position offset.
• Register the new runtime source in server/CMakeLists.txt.

Notes

• Draft note: this branch is prepared as a follow-up to feat(server): add Laguna target-layer-split adapter #297 and should be rebased once fix(server): support sampled requests in target layer split #295/feat(server): add Laguna target-layer-split adapter #297 land.
• This is intended as a follow-up to the target layer-split adapter work, not a user-facing feature change.
• The refactor keeps the adapter boundary explicit: shared runtime policy stays in common, while architecture-specific execution remains under each model directory.
• Further cleanup can move more repeated snapshot or handoff code into shared layer-split runtime helpers once the Qwen35/Gemma4/Laguna adapter shapes settle.

[cubic-dev-ai]

Review completed

_{Re-trigger cubic}