refactor(server): own per-model engine launch in model crates

[xiaguan]

What

load_engine was a per-model match that hardcoded each model's GPU topology, CUDA-Graph capability, and cross-arg validation, then built EngineLoadOptions inline — every new model fattened the server binary with model-specific knowledge.

Push that policy down into each model crate. Each crate now exposes a server-facing launch(...) that owns:

• its GPU topology defaults (V4 = 8 GPUs, V2-Lite = 2 GPUs, Qwen3 tp→device mapping, Kimi EP-world derivation),
• its CUDA-Graph capability handling (Qwen3 graph-vs-LoRA exclusion; V4/V2-Lite ignore+warn),
• its cross-arg validation (Kimi tp/dp; the old kimi_parallel_config moved here).

The server forwards CLI flags uniformly and no longer judges model capability — load_engine is now pure dispatch (≈145 lines → forwarding).

crate	new entry
qwen3-4b	Qwen3LaunchOptions + launch
qwen35-4b	launch(path, device_ordinal, cuda_graph)
kimi-k2	KimiLaunchOptions + launch
deepseek-v4	launch(path, cuda_graph, prefill_profile)
deepseek-v2-lite	launch(path, cuda_graph)

Server-crate slim

• Extract CLI Args into config.rs.
• Delete dead modules ops / tensor / weight_loader / trace_reporter and the unrun benches/ (only consumers were the benches + the deleted ops tests). sampler / scheduler kept — bench_serving uses them.
• Net ≈ -1000 LOC.

Also in this PR (separate commit)

refactor(frontend): tighten vllm-frontend serve-path visibility (lora_routes → crate-internal), extract a max-model-len helper, reshape serve_model; adapt openinfer-sim. Pre-existing working-tree work, bundled here at the author's request.

Verification

• Compiles clean (zero warnings): server+qwen3-4b (default), qwen35-4b, deepseek-v2-lite, and the bench_serving bin.
• kimi-k2 / deepseek-v4 cannot be built on the dev box — they fail inside openinfer-kernels' build.rs (missing tilelang / CUDA kernel toolchain), before any of this PR's Rust is reached. Their launch is trivial and was reviewed by inspection.
• Config arg-parsing unit tests pass (the GPU ops tests were removed with src/ops).

Deferred (not in this PR)

config.rs max_prefill_tokens default references openinfer_qwen3_4b::… without a #[cfg(feature = "qwen3-4b")] gate (its sibling LoRA fields are gated), so --no-default-features won't build. Left as-is per owner.

🤖 Generated with Claude Code