Reduce local-attention KV storage with per-layer rings

[Mikyx-1]

Closes #1016

Local-attention layers currently allocate KV storage for the entire context. This change gives each layer its own FP32 buffer: local layers use min(seq_len, window - 1 + max(1, prefill_batch_size)) rows, and global layers keep full-context storage.

The extra batch rows prevent QKV prefill writes from overwriting history needed by earlier queries. Runtime growth preserves cached history, copies remain independent, and attention reads use each layer's ring capacity while score-buffer indexing and arithmetic remain unchanged.

Validation on Gemma 3 270M and 1B: all 10 A/B comparisons produced byte-identical decoded output and identical token IDs (1,280 generated tokens per build). Runs used a shared baseline-calibrated matrix plan to isolate the cache change from timing-dependent autotuning.

Model	KV capacity, MiB	Peak process RAM, MiB	Prefill, tokens/s	Decode, tokens/s
270M	288.5 → 72.7	644.9 → 565.6	1013.3 → 1082.3	40.48 → 42.30
1B	416.5 → 100.0	1574.3 → 1458.5	199.2 → 231.7	13.79 → 15.00

Medians from three alternating pairs with a 3,367-token prompt, 128 generated tokens, 8,192-token capacity, 256-token prefill batches, and six pinned workers on an i5-12400F. CPU load varied; speed improvements are measured observations, not a universal performance guarantee. RAM savings were consistent.

• CMake cache/attention regression tests: 8/8 passed, covering wraparound, runtime growth, prefix attention, global layers, context limits, copies, and exact FP32 attention comparisons on AVX2 and EMU128.
• AddressSanitizer with leak detection: 8/8 passed. Changed KV/attention translation units and tests were instrumented; supporting libraries were reused.
• CLI, single benchmark, API server, and batch benchmark built successfully.

C++ API change: direct users of the former public KVCache::kv_cache matrix must use LayerCache(layer) or Row(layer, position). Repository callers are updated.

[Mikyx-1]

Superseded by #1020, which ports the local-cache changes to dev and includes new dev-baseline validation.