Resilient selective-testing shard downloads and bounded uploads

Selective-testing shard downloads are now resilient to transient slowdowns: the shard bundle download retries on transient failures (3x) instead of failing the run on a single hiccup, multi-GB upload bundles are split into a shared artifact plus one per .xctest module (each shard downloads only what it needs), the upload memory peak is bounded by maxConcurrentParts instead of the artifact size, and per-module xctest archives are produced in place from the products directory to avoid tens of GB of redundant temp-copy disk I/O.