Upgrade llama.cpp from b9049 to b9071

[bernardladenthin]

Summary

This PR upgrades the pinned llama.cpp version from b9049 to b9071, incorporating upstream improvements in multimodal support, attention mechanisms, backend initialization, KV state management, and hardware acceleration.

Key Changes

• Multimodal & Chat: Added contains_media() method to chat messages and improved JSON serialization for media-containing messages
• Attention Mechanisms: Introduced LLM_KV_ATTENTION_VALUE_SCALE KV key and corresponding hparam field for MiMo-V2 attention value scaling
• Backend Initialization: Fixed llama_supports_gpu_offload() and llama_supports_rpc() to auto-initialize backends if none are registered
• KV State Management: Improved state_seq_set_data by removing overly-strict seq_id matching and optimizing tensor reallocation logic to check shape compatibility
• MiMo-V2 Model: Extended with Multi-Token Prediction (MTP) layer support, fused wqkv projections, and attention value scaling
• Hardware Acceleration:
  • Added SYCL implementations for CUMSUM, DIAG, FILL, SSM_SCAN, SOLVE_TRI operations
  • Optimized CUDA outer-product to use cublasSgemmStridedBatched for batched operations
• Multimodal Tools: Added MiniCPM-V 4.6 support with new projector type and ViT merger graph
• Server UI: Integrated LLM-based conversation title generation and CSS animation fixes

Notes

All changes are additive or internal improvements with no breaking changes to the Java bindings API.

https://claude.ai/code/session_01X1BGPBMMuKRUcvRs8Su9rL