Add raw completions server mode

[Dango233]

Summary

This adds an opt-in --raw-completions mode for ds4-server.

When enabled, /v1/completions treats the request prompt as an already-rendered model prompt and continues it directly, instead of wrapping it in the built-in DS4 chat template. The default behavior is unchanged.

Why

This aligns ds4's opt-in completions behavior with llama.cpp-style raw completions, where the client owns prompt rendering and the server only performs continuation.

That is useful for clients and experiments that provide their own model-specific templates, including text-completion/instruct frontends such as SillyTavern, KoboldAI/KoboldCpp-style clients, and custom completion clients that send rendered prompts directly.

Notes

• POST /v1/chat/completions is unchanged.
• Default /v1/completions behavior is unchanged unless --raw-completions is passed.
• In raw mode, thinking markers or other model control tokens should be included by the client as part of the rendered prompt.

Testing

• make ds4-server ds4_test
• ./ds4_test --server
• ./ds4-server --help | rg -n "raw-completions|completions"