fix: preserve extra request fields when model name is remapped

[octo-patch]

Fixes #2295

Problem

When a channel has model mapping configured (e.g. {"hy-dsc": "deepseek-v3-0324"}), relay/controller/text.go was re-serialising the request from the parsed GeneralOpenAIRequest struct. Any unknown fields sent by the client — such as enable_search, enable_thinking, or other provider-specific params added via the OpenAI Python SDK's extra_body mechanism — were silently dropped in that struct → JSON round-trip, so they never reached the upstream provider.

Without the fix, a call like:

client.chat.completions.create(
    model="hy-dsc",
    messages=[...],
    extra_body={"enable_search": True},
)

would strip enable_search from the forwarded body, breaking Tencent Cloud DeepSeek's web-search feature and similar provider extensions.

Solution

For OpenAI-type channels where the only transformation needed is a model name substitution (no forced system prompt, no EnforceIncludeUsage, not Baichuan), patch the model field directly in the cached raw request body instead of going through a full parse → marshal cycle. This preserves all extra/unknown fields so they reach the upstream provider unchanged. The code falls back to the existing full-conversion path if the raw body is unavailable or unparseable.

Testing

• Configured a channel with model mapping {"alias-model": "real-model"}.
• Sent a request with model=alias-model and an extra field enable_search=true in the body.
• Verified the upstream received model=real-model and enable_search=true.
• Verified that requests without model mapping still pass through the body untouched (existing path unchanged).
• Verified that requests requiring full conversion (forced system prompt, Baichuan channels, EnforceIncludeUsage) still use the full conversion path.