PR1: per-layer KV storage (mixed-precision layer tables)

[Astrangemaninhere]

PR1: per-layer KV storage (mixed-precision layer tables)

Background

The KV cache previously used a single global storage format for every layer. Attention layers differ sharply in how much quantization error they tolerate, so a uniform format either wastes memory on robust layers (all-INT8) or degrades quality on sensitive ones (all-NVFP4). This PR makes the storage format a per-layer decision.

Changes

• Per-layer dtype table: each full-attention layer selects BF16 / INT8 / FP8 independently via --kv-layer-storage or the variant default table.
• Layer plane-base prefix sums: mixed-format layers have unequal plane counts and strides; the page geometry now tracks each layer's plane offset.
• Scale-plane resolution: FP8 per-layer scale planes are FP16 (not the global format), and per-layer quant_group is resolved into layer views.
• Table propagation: layer_kv_dtypes is copied into the sequence candidate so the table actually reaches the cache (previously never applied).

Files

• src/targets/qwen3_6/impl/state/decoder_state.cpp (plane geometry, per-layer views)
• src/targets/qwen3_6/impl/runtime/layouts_impl.h (candidate propagation)
• src/core/paged_kv_cache.h (layer views carry dtype/group/plane base)

Dependency

Foundation for PR2 (cold pool), PR5 (YaRN), and PR6 (NVFP4 tier). Merge first.