fix: improve logging parsing for empty 502 errors

[subpath]

Context

Transport-level 502s logged as Failed to proxy request: with no detail. raise_and_log (utils.py) built its message from e.response.text or str(e); httpx transport errors (ConnectError, RemoteProtocolError, timeouts) have no .response and often stringify to "", so the line was empty. Plain logger.error also captured no traceback (record.exception was null), and lines carried no user/model context.

What's new:

• raise_and_log: log type(e).name, fall back to repr(e), and use logger.opt(exception=e) to attach the traceback. Client-facing error text unchanged.
• logger.py HTTPX _wrapper: include exception type + repr (first line of the 502 triple).
• Bind request fields onto the loguru contextvar via logger.contextualize(*req.log_fields) in stream_completion/get_completion/get_search, so user/service_type/purpose/model ride in record.extra. (queryable in Cloud Logging). log_fields is a shared mixin on the chat/search request types.
• The scope is bound inside the proxy functions, not middleware: the middleware contextualize exits at the 200 OK, before the SSE body iterates, so streamed errors lose context (the known streaming blind spot). The streaming wrapper forwards aclose() to preserve disconnect handling.

QA:

Three regression tests capture real loguru records and assert message, traceback, and bound fields:

• empty-message transport error through get_completion;
• mid-stream error (verifies streaming scope);
• HTTPX wrapper line.

Verified by reverting the fix: the line collapsed back to Failed to proxy request: and both tests failed.