[ansh-rohilla]

Problem

In SpanBatcher._flush(), self._lock was held while calling _dynamic_sampling_context() and serializing spans via _to_transport_format(). This blocked concurrent additions to the buffer for the entire duration of serialization and posed potential lock-order inversion hazards with GC finalizers (similar to what was addressed in #7776 for Batcher).

Additionally, in SpanBatcher.add(), _record_lost_func was invoked while holding self._lock.

Solution

1. In SpanBatcher._flush():
   • Detach the trace buckets under self._lock via self._span_buffer.pop(bucket_id, None) and self._running_size.pop(bucket_id, None).
   • Release self._lock before constructing envelopes and serializing spans with _to_transport_format().
2. In SpanBatcher.add():
   • Move the _record_lost_func call outside self._lock when dropping spans due to queue overflow.
3. In tests/tracing/test_span_batcher.py:
   • Add regression tests verifying:
     • _to_transport_format() executes without holding _lock.
     • Concurrent additions are not blocked while spans are serialized.
     • _record_lost_func is invoked outside _lock on queue overflow.

Fixes #8023