perf(purejava): parallelize Q8 FFN preparation

[bsbodden]

Summary

• partition residual addition, RMS normalization, and Q8 activation preparation by active prompt row inside the retained seven-stage layer plan
• enable the path only for all-Q8 staged gate/up projections with block-major activations and at least two GGUF participants
• expose the default-off deployment/recommendation setting through deterministic plan schema pure-java-v14
• document the controlled full-model retention gate

Why

The post-block-major JFR profile showed stage-completion skew while one participant prepared every FFN input row. The new path balances that independent row work across the existing GGUF participants without adding a stage or changing batch-one decode.

Vectors support is merged at 523f3aa95f503a35babc02b76e5537df5f06891c.

Correctness

• test-first red/green/refactor development
• exact 40-token forward-pass comparison covering a full batch and partial tail batch
• bit-identical final logits, complete K/V buffers, and next autoregressive logits
• deterministic serial fallback for mixed formats, packed activations, and single-participant runtimes

Performance

SmolLM2 360M Q8_0, GraalVM Java 25, fixed 1 GiB heap, eight workers, batch 32, six counterbalanced fresh-process pairs, five warmups plus five measured requests per process:

Metric	Serial	Batch-row	Change
p50 TTFT	913.102 ms	899.064 ms	-1.54%
p95 TTFT	930.788 ms	914.340 ms	-1.77%
p50 prefill	172.543 tok/s	175.592 tok/s	+1.77%
p50 CPU	6,860 ms	6,840 ms	-0.29%

All six process-pair medians and 28/30 corresponding trials improved. All 60 trials succeeded with exact corresponding input counts, output counts, and output hashes; median process RSS did not regress.

Validation

• ./gradlew :models-backend-purejava:check spotlessCheck
• repository-wide Spotless configuration: all 31 checks pass