mirpy 4.8.0

Fixes #61: require vdjtools 5.3.2 and preserve caller CPU budgets for native distance and clustering operations.

• Reduce sample embedding memory by writing directly into one columnar output matrix. Apply sample PCA with bounded scratch space, reuse MMD output, and avoid sorting full pairwise matrices for neighbor-rank selection.
• Default sample workers to the allocated CPU budget, validate workspace/chunk sizes, synchronize cold model loading, and reuse IGH isotype masks.
• Add mir reference NAME: fetch a release-pinned, checksum-verified bundle from AIRR feature signatures, with offline pre-staging and forced repair. Five paired references include frozen vocabularies and matching bases; existing vdjtools models and corpus paths remain available.
• Refresh six examples and clarify documentation navigation, RFF-before-aggregation semantics, batching and native thread controls.

Validation includes Linux/macOS Python 3.11/3.12, optional Torch/ANN/BioPython/Parquet integration, installed-wheel checks, strict documentation builds, downstream TCREmp tests, and completed 10,000-sample and 100 × 1M-clonotype synthetic stress cases. On the 10,000-sample check, sample-stage peak RSS fell from 2.25 to 1.06 GiB with byte-identical output.

Full pairwise MMD output and dense DBSCAN neighborhoods can still require quadratic memory; bounded embedding does not remove those output costs. Existing fitted parameters and corpus coordinates are unchanged.