llama.cpp tests DeepSeek V3.2 support

This is infrastructure work, not hype: the hard part is making a sparse-attention MoE model behave correctly inside a local inference stack that was not originally built for it. The PR adds the lightning indexer and DSA path DeepSeek V3.2 needs, so it is about faithful model support rather than just loading weights; the testing quants are enormous, so this targets cluster-scale or very high-memory rigs; the dedicated Jinja template and tokenizer conversion caveat show the port touches model architecture, formatting, and conversion tooling, not just runtime kernels; and the CUDA OOM guidance around `ubatch` and `-fitt` suggests the branch is usable for testers but still rough around the edges. Since it is still a draft PR, the main question is correctness and maintainability upstream, not whether the branch is interesting.