Add biology_singlecell benchmarks (PBMC3k/68k) to benchmark knowledge base

[Tryboy869]

What

Adds a biology_singlecell domain entry to researchclaw/data/benchmark_knowledge.yaml,
following the existing schema (keywords / standard_benchmarks / common_baselines).

• 3 benchmarks: PBMC3k, PBMC3k-processed, PBMC68k-reduced — all loaded via
  scanpy.datasets.* (already installed in Dockerfile.biology), all tier 1
  (self-caching calls under 6MB, no separate setup-phase download needed).
• 3 baselines: Leiden clustering, K-Means, Agglomerative clustering — all using
  packages already present in Dockerfile.biology (pip: [] throughout).

Names/sizes/sample counts were checked against current scanpy documentation
rather than assumed. One thing worth a reviewer's eye: despite the name,
pbmc68k_reduced ships only a ~700-cell reduced subsample, not the full 68k
cells — flagged inline in the YAML comment so it doesn't surprise anyone later.

Why

benchmark_knowledge.yaml currently covers 16 ML/DL subfields but no non-ML
domain, even though Dockerfile.biology and the biology_* domain profiles
show biology is an actively sandboxed domain. This is a narrowly-scoped first
increment for single-cell analysis specifically (matching the existing
biology_singlecell profile). Happy to follow up with other biology
sub-domains or other domains as separate PRs per the one-concern-per-PR
guideline in CONTRIBUTING.md.

Tests

• Added TestBiologySingleCellDomain (4 tests) to tests/test_benchmark_agent.py:
  domain presence, keyword matching, tier/API-prefix correctness, and that no
  baseline needs a pip package beyond what Dockerfile.biology installs.
• pytest tests/ passes locally (2868 passed, 56 pre-existing/unrelated skips).

Scope

Additive only — no existing domains, agents, or schemas touched.