Relocate Langfuse fixtures to their own runner (#187)

* Relocate Langfuse fixtures to their own runner

The Langfuse-mapping observability fixtures are fixture-tested by the
dedicated sibling runner tests/conformance/test_observability_langfuse.py,
not by unit tests. The _UNIT_TESTED_FIXTURES bucket in test_observability.py
mislabeled them, so the fixture-harness catch-up wrongly wired
022/023/024/031/032 into test_observability.py -- duplicating coverage that
already existed in the sibling runner -- and homed the genuinely-new
035/036/059 there too.

Revert all Langfuse wiring from test_observability.py (back to its
post-tier-1 state) and add a _LANGFUSE_HARNESS_FIXTURES bucket that the
coverage guard counts and the dispatcher skips, recording that the sibling
runner owns them. Relocate 035/036/059 into test_observability_langfuse.py,
extending its runner for caller_invocation_id, the derived-trace.id bridge,
036's raw invocation_id, and the harness_parameterized / non_empty_string
metadata matcher.

Test-only; no behavior or pin change.

* Fix side-effect-in-assert in 036 trace.id check

CodeQL py/side-effect-in-assert flagged the expected_metadata.pop()
inside the assert: under python -O the assertion (and its pop) is
stripped, leaving the invocation_id key in expected_metadata for the
later _assert_metadata_subset to wrongly check. Bind the pop to a local
before the assert, and include the expected value in the failure message.

Author: chris-colinsky