Fix missing llama-server binary causing "binary not found" error

[IvanBrazza]

Summary

• Ollama 0.30+ moved inference out-of-process into a standalone llama-server binary plus supporting shared libraries (libllama-server-impl.so, libllama.so, libmtmd.so, etc.), all installed together under /usr/lib/ollama/. This Dockerfile only ever copied libggml-cpu-* and libggml-base.so, so llama-server was never present in the built image, breaking every model run with llama-server binary not found.
• Copy the whole /usr/lib/ollama directory instead of hand-picking files, and prune the GPU backend subdirs (cuda_v*, rocm_v*, vulkan, cuda_jetpack*) in a separate build stage so the deleted GPU bytes never end up in the final image's layers (a naive rm -rf in the final stage bloats the image to several GB due to layer history).

Fixes #12

Test plan

• Built the image locally and confirmed /usr/lib/ollama contains llama-server and all its dependent .so files
• Ran the container, pulled nomic-embed-text, and ran ollama run nomic-embed-text helloworld — got a real embedding back (previously failed with the binary-not-found error)
• Confirmed final image size stays close to the original ~85MB (measured ~105MB) rather than ballooning with GPU backend libraries

🤖 Generated with Claude Code