feat(observability): add bounded OTLP HTTP exporter

[PhilipJohnBasile]

Splits the OTLP/HTTP exporter out of #336.

This PR contains only the observability exporter:

• dependency-free OTLP/HTTP JSON encoding
• bounded asynchronous queue and batching
• short export timeouts and fail-open delivery
• privacy-default attribute sanitization
• explicit content opt-in
• endpoint and header configuration
• a localhost collector measurement with disabled, healthy, and dead-endpoint arms

It does not add server middleware, policy hooks, replay, memory systems, dashboard code, or workflows.

Measurement on commit 2253b9f, macOS 27.0 arm64, Python 3.14.6, 10,000 spans per arm, batch size 128:

Arm	Emit p50	Emit p95	Exported	Dropped	Request failures
Disabled	7.792 us	8.291 us	0	0	0
Healthy localhost collector	9.833 us	12.458 us	10,000	0	0
Dead endpoint	10.042 us	21.958 us	0	3,202	0

The healthy arm used a 10,000-span bounded queue, reached a high-water mark of 8,080, added 20,304 KiB to process max RSS, and leaked the sentinel content zero times. The dead-endpoint arm used a 1,024-span queue and recorded 54 failed exports without failing a caller.

Reproduce:

python scripts/bench_otlp_export.py \
  --spans 10000 \
  --batch-size 128 \
  --queue-size 1024 \
  --output /tmp/otlp-export.json

Validation:

python -m pytest tests/test_otlp_export.py
6 passed in 0.57s

python -m compileall -q mtplx/otlp_export.py tests/test_otlp_export.py scripts/bench_otlp_export.py
PASS

ruff check mtplx/otlp_export.py tests/test_otlp_export.py scripts/bench_otlp_export.py
PASS

git diff --check
PASS

This measures exporter control-path cost and localhost delivery. It makes no distributed tracing backend or live serving throughput claim.