Make applyChatTemplate a buffer-based C ABI for objcxx interop

[FuJacob]

Summary

The applyChatTemplate shipped in #6 took a std::vector<ChatMessage> (a struct with std::string members) and returned std::string. That compiles under this package's test target (Cxx interop) but does not bridge into the Cotabby app target, which uses objcxx interop — a returned std::string has no usable Swift String initializer there, so the app literally cannot call it.

This replaces it with a detokenize-style C ABI that crosses the boundary cleanly:

int applyChatTemplate(const char* system_text, const char* user_text,
                      bool add_assistant, char* buffer, int buffer_size)

system + user are passed directly (autocomplete is always exactly those two turns), matching how the app already calls tokenize/detokenize with const char* + a caller buffer. The ChatMessage struct is removed.

Return contract:

• > 0 — bytes written (<= buffer_size), the formatted prompt.
• 0 — no model, no chat template, or render failure → caller falls back to the raw prompt path.
• < 0 — -(required size); buffer too small, caller resizes and retries.

Validation

• swift build — clean.
• swift test — Executed 15 tests, 0 failures (3 model-dependent tests skipped without COTABBY_TEST_MODEL_PATH). The no-model guard test and the end-to-end chat-template assertion were updated to the new signature.

Risk / rollout notes

• API-breaking vs Add chat-template rendering and special-token tokenize path #6, but Add chat-template rendering and special-token tokenize path #6's applyChatTemplate/ChatMessage had no working consumer (the app couldn't call them), so nothing is actually broken. hasChatTemplate and tokenizeWithOptions are unchanged.
• The Cotabby app PR that consumes this is opening alongside.