fix: drop None attributes from OTel span events

[lesnik512]

_emit_event now leaves out None-valued attributes from the OTel span event. The log record still gets every attribute, None included.

Why

OpenTelemetry SDKs before 1.45 reject None attribute values. I reproduced it with an in-memory exporter, a recording span, and a sync Retry(max_attempts=2) that gives up on a raised ConnectError through MockTransport. On SDK 1.20.0, 1.38.0 and 1.44.0 every give-up logged

WARNING opentelemetry.attributes: Invalid type NoneType for attribute 'last_status' value. Expected one of ['bool', 'str', 'bytes', 'int', 'float'] or a sequence of those types

and the recorded retry.giving_up event had no last_status. SDK 1.45.1 accepts None and keeps it without a warning. Our floor is opentelemetry-api>=1.20, so most of the supported range is affected. With this change, every SDK version records the same event (without the None keys) and none of them log the warning. On 1.45+ that means last_status: None no longer shows up on the span event, so a missing key now means "no value".

Events that can carry None:

• retry.giving_up (reason="max_attempts"): last_status when the last attempt raised instead of returning a response.
• retry.budget_refused: last_status, same case.
• circuit.rejected: retry_after when the circuit is half-open with a probe in flight.

bulkhead.rejected declares acquire_timeout: float | None, but it is only emitted when a bounded acquire times out, so it is never None in practice. The retry_after_exceeds_max_delay give-up always has a response.

The test follows the existing mock-span pattern in tests/test_observability.py, so no opentelemetry-sdk dev dependency was added.

Performance

_emit_event runs only when a resilience event fires, not on every request. The filter sits inside the OTel branch, so log-only installs don't pay for it. A dict comprehension over a six-key attribute dict costs about 0.6 to 0.7 us. Timing the whole _emit_event against a recording SDK span came out at 28 to 44 us per call before and after, and the run-to-run noise was bigger than the difference. On SDK 1.44 the None case got faster, from 42 to 54 us down to about 30 us, because the SDK warning is no longer logged.

Closes #171