fix(serving): prefix-cache binding, hit stats, dtype (needs #197)

[drunkcoding]

Productizing prefix-kv-reuse. Depends on #197 (fix/expert-dispatch-stream-ordering → dev): without it, prefix-enabled runs are corrupted by an expert-dispatch stream race.

Fixes in this PR

• kv_cache.set_block_store: normalize the cuda device index so cuda vs cuda:0 geometry no longer rejects the store — the prefix cache now actually binds/activates.
• prefix_cache + engine: surface prefix_cache_hits_total / prefix_cache_matched_tokens_total.
• attention_backend: pass q_data_type/kv_data_type to the FlashInfer plans.
• big_modeling: eager attention for qwen3 when flash-attn is unusable (sm120).
• prefix_cache_benchmark: subprocess-per-mode; --parity-report (tensor dumps + numerical analysis), --device-memory-ratio/--kv-cache-ratio for shared GPUs.
• tests/integration/test_flashinfer_kernel_parity.py: full-prefill and append kernel schedules are each deterministic, within ULP-scale error of an fp32 reference, and mask partial-page tails (3/3 pass on sm120).
• tests/serving/test_prefix_cache_cuda.py: parity test now runs one-engine-per-subprocess (kTopologyHandle-safe) with sound invariants.

HOLD resolved — what "warm not bit-exact on sm120" actually was

1. FlashInfer kernels: exonerated. Cross-kernel bitwise equality was never an invariant (per-op delta ≈ 1 bf16 ULP; both paths equidistant from fp32 truth; partial-page masking correct). See the committed kernel-parity test.
2. Real bug: expert-dispatch stream race (pre-existing on dev, exposed by prefix-mode timing) — found via NaN-poisoning: enabled_cold produced 100%-NaN logits from layers[0].mlp with finite inputs. Fixed in fix(dispatcher): stream-order expert exec against input producer and consumers #197.
3. Post-fix validation (sm120, Qwen3-30B-A3B, 64-token shared prefix + 8-token suffix):
   • disabled == enabled_cold bitwise (token digest e37a9d14, logits Δ = 0.0).
   • Warm reuse active (geometry q=8, ctx=64, kv=72, hit counted). Warm step-0 argmax swaps within the shared top-2 pair (cold: 2157@22.25 / 85355@21.5; warm: 85355@22.125 / 2157@21.5) — near-tie amplification of ULP-level kernel deltas through 48 MoE layers.
   • Token-level determinism is not a property of the engine: same-mode warm reruns flip step-1 tokens because expert-output accumulation order (final_hidden_states_.add_ / index_add_) is completion-order nondeterministic (pre-existing, mode-independent). The historical warm digest 3739c3b3 reproduces exactly this way.

Merge order

dev ← #197 first, then #181 (rebased on dev), then this PR into #181.