ec2 build: narrow CUDA kernel compilation by quant type and GPU arch

[kmbandy]

Problem

llama.cpp builds CUDA kernels for every quantization type × attention pattern combination, even when we're only using one or two quant types. This results in 100+ .cu.o compilations and 20-30 min cold builds on EC2.

Proposed Fix

Add a QUANT_TYPES variable to ec2-tool-calls-gen.sh (and related gen scripts) that limits template instantiation at build time:

# e.g. only build kernels needed for iq4_nl
QUANT_TYPES="iq4_nl"

Translate this into the appropriate CMake flags before the build step — likely via -DGGML_CUDA_FA_ALL_QUANTS=OFF and targeted -DCMAKE_CUDA_FLAGS overrides, or by patching the template instance CMakeLists before compilation.

Combined with the GPU arch already auto-detected (sm${GPU_ARCH}), this should cut cold build times significantly for single-quant jobs.

Notes

• llama.cpp has some coarse flags (GGML_CUDA_FA_ALL_QUANTS etc.) but nothing fine-grained out of the box
• The S3 build cache key would need to incorporate quant types: sm${GPU_ARCH}_cuda${CUDA_VER}_${QUANT_TYPES}
• Apply to both ec2-tool-calls-gen.sh and ec2-gpu-expert-gen.sh