[deadprogram]

Stacked on #438. Merge #437 and #438 first.

ggml-org/llama.cpp#29818 stores the decision type, temperatures and a systemone prompt template in the GGUF. decide.Open reads them, so the ggml-org OpenJev and Lev GGUFs need no config file.

• The prompt is the GGUF's systemone template, rendered with ardanlabs/jinja. JSON values are written first as the llama.cpp tojson filter writes them, with the same separators, key order and float format.
• OpenJev reads the letters A to Z and a to z at the last token.
• Lev reads A to Z and then AA to ZZ, reads choice questions in both orders and averages them, sorts JSON keys, and reads noul as a rating from 0 to 8.
• Temperatures use the per-type buckets from the GGUF.
• ParseRequest now writes non-string instructions and descriptions in the same tojson format.
• The decide example gets -readout gguf, which needs no -config.

Tested with

• go test ./exp/decide, and the wasm unit tests
• TestLevModel with lev-Q8_0, exact and batched. Decide matches DecideMany exactly.
• llama-server built at the merge commit of #29818, on the request from that PR. Every Lev probability matches within 6e-9, and the input tokens are the same (1194). It also runs on b11146.
• TinyGo build of examples/wasm/decide, which grows by 38 KB.

OpenJev is covered by unit tests against its template only. The model is 19 GB at Q4_K_M and was not run.