Per-shard test-product downloads with resilient shard transfer

Selective-testing shard downloads are now resilient (3x retry) and per-shard. The build job uploads one shared artifact plus one per-module .xctest artifact, and each shard downloads only the shared bundle plus its own modules, cutting per-shard egress around 19% on real plans. Uploads are bounded by concurrent in-flight parts so multi-GB bundles no longer OOM, and modules are archived in place to avoid copying every test bundle to a temp dir before compression.