docs(methods): consolidate 5 methods pages to 2 (preserve WASP2-unique contracts)

[Jaureguy760]

Summary

Follow-up to the user's "delete the stats stuff" request. After running a `/co-scientist` hypothesis-critic and a `/sophia-claudette` decision-auditor on the blanket-deletion proposal, both agents independently identified 4 WASP2-unique contracts in the methods pages that are not in any paper or textbook and would be silently lost by full deletion. The agreed reframe: consolidate instead of erase.

What changed

Before (5 methods pages, 883 lines):

• counting_algorithm.rst (193)
• mapping_filter.rst (329)
• statistical_models.rst (173)
• dispersion_estimation.rst (126)
• fdr_correction.rst (66)
• methods/index.rst (49, had the landing + 3 citations, 2 unused)

After (2 methods pages, ~500 lines):

• mapping_filter.rst (~335 — added counting-side canonical-filter paragraph)
• statistical_models.rst (~180 — absorbs dispersion + FDR + defaults)

WASP2-unique contracts preserved

Four load-bearing pieces flagged by the review agents, all preserved:

1. Canonical filter contract at count step — only `0x4` unmapped filter; no secondary/supplementary/dup/QC-fail filtering at count time. Not in van de Geijn 2015. Previously at `counting_algorithm.rst:69-75`, now a paragraph in `mapping_filter.rst`.
2. Profile-LRT ρ convention — ρ held at its null-model MLE when maximizing L₁; WASP2 does NOT jointly re-estimate. This was the scientific-correctness fix just shipped in PR docs: reduce text volume (~47%) + fix scientific correctness issues #99. Preserved as an `.. important::` admonition in `statistical_models.rst`.
3. BH-NaN propagation warning — `np.minimum.accumulate` silently propagates NaN through cumulative minima; `scipy.stats.false_discovery_control` raises. Matches our own MEMORY.md pitfall. Preserved as a `.. warning::` in `statistical_models.rst`.
4. WASP2 defaults and bounds — pseudocount=1, min_count=10, ρ ∈ (1e-6, 1-1e-6), logit clip ±10. Preserved as a parameter table in `statistical_models.rst`.

What got cut

• Generic biological-rationale prose ("ASE arises from...", FWER/FDR definitions, binomial-vs-beta-binomial motivation)
• CIGAR-operations reference table (standard SAM spec — trivially findable)
• Q-score formula `Q = -10 log10(P_error)` (textbook)
• Unused Storey q-value code block (WASP2 uses BH, not Storey)
• `methods/index.rst` landing page (just a captioned toctree + duplicate citations)

Cross-references updated

• `faq.rst`: `methods/fdr_correction` → `methods/statistical_models` (the NaN warning lives there now)
• `tutorials/bulk_workflow.rst`: removed duplicate `fdr_correction` ref from the FDR paragraph
• Top-level `index.rst` toctree: 6 methods entries → 2

Side benefit: fixes pre-existing duplicate-citation warnings

Removing `methods/index.rst` and deleting `dispersion_estimation.rst` eliminates the 3 duplicate-citation warnings (Kumasaka2016, vandeGeijn2015) that shipped with Tier-B.

Test plan

• `sphinx-build -q` completes with zero new warnings
• Zero broken `:doc:` refs (grep-verified across all docs)
• No Python changes → ruff/pytest/Rust tests are no-ops (but CI will run them)
• Master CI on merge commit

Cumulative reduction (across Tier-B + Tier-D + Tier-E)

• docs/source: ~6,300 → ~2,680 lines (−57%)
• Methods section: 1,357 → ~500 lines (−63%)
• Tutorial section: 3,495 → 520 lines (−85%)
• Total Sphinx pages: 30 → 20