fix(otel): fix concurrency bug in TracingHook span tracking

[simonvdk-mistral]

Bug

TracingHook stored the active span as self.request_span — a single instance attribute shared across all concurrent requests. When two requests were in-flight simultaneously, they clobbered each other's span:

Request A                          Request B
─────────                          ─────────
1. before_request()
   self.request_span = span_A
                                   2. before_request()
                                      self.request_span = span_B  ← overwrites span_A
3. after_success()
   uses span_B instead of span_A  ← wrong span
   self.request_span = None
                                   4. after_success()
                                      self.request_span is None   ← span lost

Impact: spans attributed to wrong requests, spans never closed (leaked), response attributes applied to wrong traces.

Reproducing

The test test_concurrent_async_requests_get_correct_spans reproduces this using a real Mistral client with asyncio.gather and a mock transport. An asyncio.Event gate ensures both before_request hooks fire before either response returns, guaranteeing the interleaving.

Fix

Store the span on request.extensions["_tracing_span"] instead of self.request_span. Retrieve it in after_success/after_error via response.request.extensions. This is the same extensions dict that httpx uses internally (e.g. for timeouts) — it is purely in-memory metadata, never serialized or sent over the wire.

Note: no other hook had this bug — TracingHook was the only one storing per-request state on self.