perf(purejava): select row-accumulated Q8 kernel

[bsbodden]

What

• add models.purejava.q8BlockMajorRowAccumulator to validated model/deployment settings
• select the typed Vectors Q8 block-major kernel in pure-Java plan schema pure-java-v15
• enable row accumulation only for requested Graal JVMCI staged block-major Q8 plans
• pass the selected kernel through the transformer-owned staged layer schedule
• verify exact logits, complete KV state, and the next autoregressive step
• record the controlled kernel and full-model evidence

Why

The established Q8 block-major kernel updates a widely separated output cell after every weight block. The new Vectors strategy retains one batch-sized accumulator per active output row and scatters that row once. It is a strong win on the measured GraalVM/EPYC prefill profile but neutral on local HotSpot C2, so Models must own the runtime/topology policy rather than changing the generic Vectors default.

Selection

ROW_ACCUMULATED requires all of the following:

• explicit recommendation or deployment request
• Graal JVMCI
• staged block-major Q8 prefill with an eligible Q8 consumer

Otherwise the immutable plan records the reason and selects SCATTERED.

Evidence

• controlled EPYC/GraalVM 1024x2048 batch-32 JMH: 20.571 ms/op to 4.335 ms/op (-78.9%)
• 30-trial exact SmolLM2 360M Q8_0 gate:
  • p50 TTFT: 898.794 ms to 878.801 ms (-2.22%)
  • p95 TTFT: 919.981 ms to 886.209 ms (-3.67%)
  • p50 prefill: 175.219 tok/s to 179.479 tok/s (+2.43%)
  • p50 CPU time: 6840 ms to 6680 ms (-2.34%)
  • every corresponding token count and output hash matched
• local Java 25 HotSpot/C2 JMH was neutral, so it is not selected there

Checks

• ./gradlew :models-backend-purejava:check
• ./gradlew spotlessCheck