feat: sync llama.cpp

[jhen0409]

sync to 0ca416c

[jhen0409]

ggerganov/llama.cpp#10906 seems to break builds on linux arm64 platform, I need to figure out later.

[jhen0409]

Windows ARM64 build is broken: https://github.com/mybigday/llama.node/actions/runs/12441290233/job/34737975818#step:9:1287

@hans00 any ideas?

[hans00]

Seems clang-cl also not work.
But MinGW chould not work with delayload toolchain (Node API linking).

Maybe need #15 or patch CMakeLists.txt to make it buildable.

P.S. MSVC looks not support FP16

[hans00]

patch CMakeLists.txt

Looks this change could build.

---             message(FATAL_ERROR "MSVC is not supported for ARM, use clang")
+++             list(APPEND ARCH_FLAGS /arch:armv8.7)

P.S. X Elite is ARMv8.7-A

[jhen0409]

patch CMakeLists.txt

Looks this change could build.

---             message(FATAL_ERROR "MSVC is not supported for ARM, use clang")
+++             list(APPEND ARCH_FLAGS /arch:armv8.7)

P.S. X Elite is ARMv8.7-A

Confirmed it works, but latest commit of llama.cpp still broken on linux arm64 build. I probably don't have time to deal with that, so I just rollback to ggerganov/llama.cpp@0a11f8b.