perf: Add vector search stage metrics

[franciscojavierarceo]

What changed

• Adds ogx.vector_io.query_stage_duration_seconds for retrieval stage timing.
• Adds ogx.vector_io.query_result_count for returned chunk counts.
• Emits stage metrics from the shared vector store query path for embedding generation, backend search, and neural reranking.
• Adds unit coverage for metric instrument definitions and stage emission from VectorStoreWithIndex.query_chunks.

Why

The existing router-level retrieval duration tells us that search was slow, but not whether the regression came from embedding, backend lookup, reranking, or unexpectedly large result sets. These stage-level metrics make future RAG/search regressions easier to localize.

Review notes

This draft intentionally includes the same hot-path logging cleanup as #5972 because both changes touch src/ogx/providers/utils/memory/vector_store.py and the logging hook inspects that file. If #5972 lands first, this PR should shrink to the metric changes after updating.

Validation

• git diff --check
• Commit hooks passed, including ruff, ruff format, mypy, provider codegen, logging checks, and repository policy hooks.
• Attempted targeted test: uv run pytest tests/unit/telemetry/test_vector_io_metrics.py tests/unit/rag/test_vector_store.py -q, but direct uv run did not start because the shell uv is 0.5.29 and the repo requires >=0.7.0.