Restore analysis parity and correct integrated workflow behavior

[Jaureguy760]

Single-cell mixed phased/unphased genotypes could select phased inference and crash, ATAC workflow preprocessing discarded peak membership, and fragment coverage was passed into allelic inference without observed alleles. This PR combines #138 and #139 with the remaining integration fixes and repository cleanup.

Changes:

• Restore archived genotype-set phasing, require Rust for normal single-cell counting, and retain explicitly invoked Python reference counting for validation.
• Check bulk count, pseudocount, threshold and regional-total arithmetic; reject unsupported UInt32 totals instead of wrapping.
• Restore RNA annotation defaults and parent grouping; preserve GTF inputs, region membership and zero-valued analysis parameters in workflows.
• Represent fragment-only scATAC outputs as coverage and prevent allele-specific inference from them. BAM workflow downstream inference remains unwired; the single-cell inference CLI remains Python/SciPy.
• Run five actual Nextflow process regressions and one workflow-boundary check in required CI. Repair single-cell pipeline test discovery so its compatibility check executes.
• Correct tutorials, input/output schemas, backend guidance and development installation. Remove obsolete audit/design/performance reports and seven empty Rust test placeholders; retain substantive regressions and fixtures.
• Align Rust 1.88 requirements and environment locks, declare CMake for source builds, check the Cargo root version, and replace broken SIF setup with pinned Apptainer installation.

Validation:

• Final local source: 816 Python tests passed, 82 skipped, one deselected; 134 Rust tests passed, zero ignored unit tests. Three existing documentation examples remain ignored.
• Clippy with warnings denied, Ruff, formatting, release-version checks and Sphinx with warnings as errors pass.
• All six focused workflow regressions pass with Nextflow 24.10.4/nf-test 0.9.2. Each of the four pipeline compatibility commands executes one passing test.
• Public 10x GM12878/EL4 chromosome 21 data: exact archived/current Python/Rust counting, including a final-wheel recount at threads 1/4; 54/54 widened-reference single-cell statistical configurations, 24/24 CLI comparisons, and 16/16 widened bulk CSV configurations pass.
• Untouched narrow legacy references retain their numerical failures; the tests do not relabel them as parity. Statistical models, first-SNP anchoring, chi-square df=1 and BH procedures remain unchanged.
• macOS arm64 wheel and isolated source-distribution installation succeed; all three installed entry points run outside the checkout. Source artifact inspection found no investigation files or build caches.

Limits: real-data validation covers one chromosome and donor genotype on provider-aligned BAMs, with technical cell partitions. It does not certify biological calibration, a new remapping pass, complete BAM-based single-cell Nextflow inference, or a full release matrix. All seven final-head PR workflows passed: core CI (Python 3.10–3.13, Rust, Clippy and lint), Documentation, CodeQL, Security, Linux/macOS wheel Build, Ecosystem (Galaxy, four pipeline compatibility checks and six native process/boundary regressions), and native amd64/arm64 Container CI. Linux/Python 3.11 reports 822 passed, 76 skipped, one deselected. No release is tagged or published.

Supersedes #138 and #139; close those PRs after this replacement merges. Investigation inputs, scripts and reports remain outside Git. Final reviewed head: 061e7e7; source tree: 8e67178bfc1ad4dbc2a162d834fecbccf3f49bda. Runtime source matches locally tested commit efee99a; the follow-up adds only CI dependencies needed by the now-active single-cell H5AD stub. Local and remote final trees match.