feat: sync llama.cpp to b10054

[github-actions]

🤖 Automated llama.cpp sync

This PR was automatically created/updated by the daily sync workflow.

Changes:

• Updated llama.cpp submodule from b9982 to b10054
• Regenerated bindings and build files

Verification:

• ✅ Bootstrap script completed successfully (including iOS Metal compilation)
• ✅ iOS frameworks build completed successfully (macOS job)
• ✅ Android libraries build completed successfully (Ubuntu job)
• ✅ C++ unit tests passed
• ✅ TypeScript build completed successfully

📋 llama.cpp changes (b9982 → b10054)

• ac2557c docs: added a note about using OpenCl with Adreno 810 (#25786)
• 0dc74e3 DeepseekV4: Add fused hyper-connection ops (#25585)
• b2dd28a hexagon: L2 cache handling rework (dirty bit tracking with lazy flushing) and more MUL_MAT updates (#25762)
• f15bd60 kleidiai: Add SME vs SME2 distinction in kernel dispatch (#25478)
• b15ca93 vulkan: when using transfer queue for async copies, sync on event_wait to avoid race (#25229)
• 3278e92 conversion: accept BitNetForCausalLM architecture name (#25769)
• 2e1fd76 TP: fix Phi3, Bert, Plamo2/3, ChatGLM (#25536)
• 86b719b vendor: update BoringSSL to 0.20260713.0 (#25624)
• 32e789f tests: actually exercise test-recurrent-state-rollback (#25758)
• a8dc0e3 server : allow text-only slot save/restore with mtmd (#25076)
• a55a8c5 convert : fix dflash target tokenizer mismatch during conversion (#25733)
• 79bba02 CUDA: Support CUDA Virtual Devices (#25228)
• 3f08ef2 Enable CUDA graphs on volta+turing (#25749)
• 8ee54c8 server: Ignore empty / non-existing Origin headers (#25756)
• c7d8722 ggml-cuda : restore prop.integrated on HIP builds (#24233)
• 5839ba3 CUDA: dedup MoE gate/up activation quantization (#25441)
• a320cbf ci : add official website link to release notes (#25728)
• 56d6e9d quant : allow using manual tensor types with --pure (#25716)
• 3dafb58 opencl: disable FA and MoE weights repack to work around compiler issues for Adreno 850 GPU (#25745)
• 602f828 cuda: extract Q1_0 elements via __byte_perm (#25628)
• 505b1ed opencl: exclude some moe kernels on Adreno a7x (#25698)
• 32beb24 ui: Agentic Content UX improvements (#25450)
• 3b53219 cuda : CUDA GGML_OP_LIGHTNING_INDEXER implementation (generic vector kernel + wmma kernel) (#25545)
• aff6eb6 tokenize : drop --stdin mutual-exclusion check (#25672)
• c3d47e6 opencl: fix two issues on flash attention for Adreno a7x (#25697)
• f6f12e4 CUDA: tighter MMQ src1 buffer size for native fp4 (#25613)
• 956973c Fix crash with draft-simple (#25720)
• a582222 server: fix read_file append_loc space breaking edit_file match (#25705)
• a05df0a ui: fix thinking menu never appearing in single-model mode (#25637)
• a3e5b96 cuda : relax tensor contiguity requirements for quantized concat (#25678)
• c810293 ci : add HF_TOKEN to self-hosted workflows (#25706)
• b3c9d1b metal: fuse snake activation (mul, sin, sqr, mul, add) (#25459)
• f955e39 ggml: add f16 out_prod support for CPU and out_prod op for Vulkan (#23997)
• 33a75f4 DeepseekV4: reduce graph splits (#25702)
• d3fba0c sycl : fix get_rows Q2_K, Q4_K, Q5_K (#25656)
• ae9291e sycl : support kernel type fp16 for conv2d_dw (#25653)
• 22b208b sycl : implement xielu op (#25550)
• 0e148a5 sycl: Increase minimum buffer size for USM system allocations (#25525)
• 32b741c [SYCL] Flash Attention with XMX engine via oneDNN (#25222)
• 12127de opencl: do not use clCreateBufferWithProperties when targeting CL 2.x (#25673)
• 00fa7cb opencl: handle OOB write in noshuffle GEMV kernels (odd ne01) (#25640)
• a4ce259 opencl: avoid the vec path in GEMV for unaligned row stride (#25671)
• c718542 hexagon: fix hmx-queue signal enum-narrowing problem (#25677)
• bf2c86d server : refactor prompt cache state ownership (#25649)
• 6e52db5 server: add --cors-* options (#25655)
• 236ab57 ui: Fix spacing in tool-call request (#25634)
• dfba90d webui: parse effective-parameter sizes (E2B, E4B) as params (#25529)
• 00e79f6 opencl: fix a dp4a bug for devices where cl_khr_integer_dot_product is unavailable (#25639)
• 17a05e4 ui: fix mcp panel for toggle + timeout + proxy + ON/OFF state (#25631)
• 7f575c3 DeepseekV4: fix seq_rm (#25588)

Please review and merge if all checks pass.