Switch to C++17

llama.cpp [just adopted it](https://github.com/ggerganov/llama.cpp/pull/10570), and that's also what [gemma.cpp](https://github.com/google/gemma.cpp) uses.

Some options:
- [x] Retire nonstd_make_unique backfill
- [x] Destructure pairs
- [ ] std::optional for attempted parsing
- [ ] std::variant for Value?
- [ ] More elegant function / filter definitions (leverage new auto inference)
- [ ] (in the process, optionally make the dep to json a bit looser to be able to switch to [json.cpp](https://github.com/jart/json.cpp))