feat(concurrency): concurrent paged serving

[Graffioh]

Dependency

Builds on #555, which provides the paged K/V allocator, paged-attention operator, and single-request Qwen paged-decode foundation.

Summary

Adds concurrent paged serving for Qwen behind the existing paged-attention path.

• Introduces model-neutral SeqEngine, StepPlan, and StepResult contracts.
• Adds scheduler-owned admission, fair automatic prefill distribution, request retirement, cancellation handling, and slow-client buffering.
• Derives idle and mixed prefill capacity from engine-advertised limits instead of exposing scheduler-internal token-budget flags.
• Batches live Qwen decode rows and supports fused prefill/decode lowering.
• Stores recurrent state per sequence with explicit row-to-slot mapping.
• Reserves prompt capacity atomically and derives paged-K/V capacity from available memory.
• Adds host, kernel, contract, and end-to-end parallel-serving coverage.
• Extends CI to exercise concurrent-serving kernels while retaining current DeepSeek4 coverage.

Impact

Multiple requests can share one Qwen backend without run-to-completion serialization. Slot lifecycle, paged K/V ownership, recurrent state, sampling state, and response delivery remain isolated per request.

The normal serving interface stays small: users select maximum concurrency and may override K/V pool capacity, while the engine supplies appropriate idle and mixed prefill limits automatically.

Validation

• git diff --check
• test_seq_batch_plan: 41 checks passed
• Full HIP test_server_unit target built successfully for gfx1151
• Focused server configuration and /props tests passed
• CUDA configuration from current main
• Focused CUDA build progressed through the modified core and concurrent-serving kernel sources without errors; the full GPU template build was not completed locally

TODO future

• shared paged prefixes

[cubic-dev-ai]

All reported issues were addressed across 64 files

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

Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 67 files

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

Re-trigger cubic</sub>