Add chunked Qwen prefill with backend KV cache reuse

[0xShug0]

Summary

Extract reusable framework support for bounded, single-sequence Qwen prefill directly into the backend KV cache.

Additions

• Add QwenCausalDecodeRuntime::prefill_embeddings_into_cache() to process embedding inputs in bounded blocks and continue token decoding from the same cache, without exporting and reimporting KV state through host memory.
• Add multi-token cache updates through FastKVSetRowsModule::build_block() and QwenDecoderLayerModule::build_with_static_cache_block().
• Add TransformerKVCache::clear_on_backend() using the existing ggml tensor-memset API. It zeroes key/value tensors and resets position counters while retaining their allocations.
• Keep the block-prefill graph owned and reused by the runtime, with cleanup integrated into its existing graph lifecycle.
• Add a focused unit test and CMake registration for chunked prefill, continued decoding, repeated use, cache clearing, and capacity handling.

Scope

The new prefill path is explicit opt-in and requires DirectSetRows with full last-token readback. Existing prefill entry points remain available. This PR contains no model integrations, model specifications, UI changes, or ggml modifications.