fix: update cpu_get_num_math to common_cpu_get_num_math (llama.cpp rename)

[WayOfTheMap]

Summary

After npx node-llama-cpp source download --release latest --gpu metal, the addon fails to compile with:

addon/AddonContext.cpp:400:41: error: use of undeclared identifier 'cpu_get_num_math'
addon/AddonContext.cpp:748:75: error: use of undeclared identifier 'cpu_get_num_math'
addon/addon.cpp:54:42:        error: use of undeclared identifier 'cpu_get_num_math'

Upstream llama.cpp renamed cpu_get_num_math() to common_cpu_get_num_math() in common/common.h as part of the common_* namespace pass. The function signature is unchanged.

This PR updates the three call sites in the addon to use the new name.

Changes

• llama/addon/AddonContext.cpp:400 — cpu_get_num_math() → common_cpu_get_num_math()
• llama/addon/AddonContext.cpp:748 — cpu_get_num_math() → common_cpu_get_num_math()
• llama/addon/addon.cpp:54 — cpu_get_num_math() → common_cpu_get_num_math()

Related

This is one of three compile blockers when rebuilding the addon against current llama.cpp on macOS Sequoia (SDK 26.2). The other two — the common → llama-common library target rename and the std::atomic_bool loaded = false; deleted-copy-ctor — are addressed by #597 (@CreatiCoding). All three patches together produce a working build.

Verification

Applied this patch plus #597 against withcatai/node-llama-cpp@v3.18.1, ran npx node-llama-cpp source download --release latest --gpu metal (resolved to llama.cpp b9145), and verified end-to-end on macOS 15.x arm64 with Metal:

Model	Load	Generate
Llama 3.2 3B Instruct Q4_K_M	✓	✓
Qwen 2.5 7B Instruct Q4_K_M	✓	✓
Gemma 4 E4B Instruct Q4_K_M	✓	✓

Gemma 4 specifically requires this rebuild path — the prebuilt llama.cpp b8390 that ships with v3.18.1 predates Gemma 4 architecture support, so loading a Gemma 4 GGUF without rebuilding returns unknown model architecture: 'gemma4'.

[andreinknv]

+1 — independently hit the same three undeclared-identifier errors today while rebuilding the addon against current llama.cpp on macOS arm64 (b9151). This patch fixes them cleanly; combined with #597 (the common → llama-common link target rename + the atomic_bool loaded = false brace-init) the addon builds end-to-end. Verified locally with Qwen2.5-Coder Q4_K_M, jina-embeddings-v2-base-code, and bge-reranker-v2-m3 — all load + decode correctly on Metal.

Cleanly extractable from a larger working tree; great PR scope.