fix(pipeline): pump _stream_with_keepalive source from one task, not one per item

[sjawhar]

Problem

opentelemetry.context logs Failed to detach context (ERROR, with a full
traceback) roughly twice per streaming request. In production this produced
~60k ERROR lines per 48h, continuously, since the streaming-keepalive change
shipped (#804) — noise that also feeds a Datadog error-rate monitor (which
now excludes this logger by attribute purely because of the volume) and,
separately, exhausted the Sentry error quota when ignore_logger wasn't
applied to it (#807).

Root cause

AnthropicClient.stream() and _AnthropicPolicyIO._stream() each hold an
OpenTelemetry span open across every chunk of the upstream response:

with tracer.start_as_current_span("anthropic.stream"):
    async for event in stream:
        yield event

_stream_with_keepalive (introduced in #804) drove that generator by
wrapping every single __anext__() call in a fresh
asyncio.ensure_future(...) Task:

if pending is None:
    pending = asyncio.ensure_future(_anext_or_done(iterator))
item = await asyncio.wait_for(asyncio.shield(pending), interval_seconds)
pending = None  # next item gets a brand-new Task

asyncio.Task copies contextvars.Context at creation. The span's
context-attach token is created in whichever Task fetches the first chunk;
the span's __exit__ (on the last chunk / StopAsyncIteration) runs in
whichever Task fetches the last chunk — a different Context object.
contextvars.Context.reset() raises ValueError in that case, and
opentelemetry.context.detach() catches it and logs it at ERROR rather than
propagating it — hence the silent-but-noisy storm (the request itself
completes with a 200; otel-collector shows no export failures).

Reproduced directly against _stream_with_keepalive with a real
TracerProvider, no Sentry involved — same ValueError and the same
opentelemetry/context/__init__.py:157 traceback line seen in production.
Sentry enablement was the leading hypothesis (present in prod, absent in a
Sentry-less staging comparison), but isn't the mechanism: our init_sentry()
doesn't set _experiments.otel_powered_performance, so Sentry's
OpenTelemetryIntegration (which would install its own OTel context/span
processor) never activates, and Sentry's AsyncioIntegration task factory
wraps coroutines without touching opentelemetry.context at all. The storm
is asyncio/OTel-only and would reproduce on any environment carrying real
stream=True traffic once #804 is deployed, independent of Sentry.

Fix

_stream_with_keepalive now drives source from a single persistent
background task (_pump_to_queue) that feeds a one-slot asyncio.Queue,
instead of a fresh task per item. This keeps every span's attach/detach pair
inside that one task's contextvars.Context for the generator's entire
lifetime, no matter how many keepalive timeouts happen in between.

Preserves the existing contract:

• A slow item is never dropped (the pump keeps running independently of the
  consumer's per-iteration timeout).
• Closing the consumer (aclose()) cancels the in-flight pump and propagates
  cancellation into source.
• Exceptions raised by source (including a spontaneous CancelledError, as
  opposed to this task being cancelled from the outside) propagate to the
  consumer instead of hanging.

Depends on

This branch is based on #804 (fix/anthropic-stream-keepalive), which
introduced _stream_with_keepalive/_anext_or_done. It should land after
#804 merges; the diff below includes #804's commits until then.

Verification

• New regression test
  (TestStreamWithKeepalive::test_span_spanning_multiple_yields_detaches_cleanly)
  reproduces the exact Failed to detach context ERROR against the pre-fix
  code and passes after the fix.
• New test (test_source_exception_propagates) covers exception relay
  through the new queue-based pump.
• uv run pytest tests/luthien_proxy/unit_tests/pipeline/test_anthropic_processor.py — all pass.
• uv run pytest tests/luthien_proxy/unit_tests (full unit suite) — all pass.
• uv run ruff check / uv run ruff format --check — clean on touched files.
• uv run pyright — 0 errors, 0 warnings, 0 informations.
• ./scripts/dev_checks.sh --fast — same 3 pre-existing, environment-order
  failures already called out in fix(sentry): never capture opentelemetry.context detach noise #807's verification notes
  (test_onboard.py::test_find_docker_ports_respects_env_vars and two in
  test_config_registry.py), each confirmed passing in isolation; unrelated
  to this change.

Changelog fragment included.