CompressedKvAttention

Bridge between KvCacheStore and the SDPA execution path.

This abstraction provides the integration point for compressed K/V in the attention runtime. Instead of modifying the core TensorOps interface (which maps to backend-specific fused kernels), this component sits between the model layer and SDPA:

1. Write path: Compresses K/V on token append via storeKeyValue

2. Read path: Dequantizes only required tiles via loadKeysForAttention and loadValuesForAttention

3. Extension point: Backends can override DequantStrategy to fuse decompression with attention math.

Usage in a transformer layer:

val bridge = CompressedKvAttention(kvCache)
bridge.storeKeyValue(layer, keyProjection, valueProjection)
val keys = bridge.loadKeysForAttention(layer)
val values = bridge.loadValuesForAttention(layer)
// pass keys, values to scaledDotProductAttention