Selective-testing shard downloads retry and download only the modules each shard needs

Sharded test downloads now retry (3×) on transient slowdowns instead of failing on the 90s timeoutIntervalForResource cap, and the shard bundle is split into one shared artifact (frameworks, dylibs, .xctestrun) plus one per-module .xctest artifact. Each shard downloads only what it needs, the suite bin-packer now groups a module’s suites onto one shard (oversized modules still split) for the “where modules are reused across shards” benefit, and very large uploads are bounded to maxConcurrentParts * partSize instead of streaming whole multi-GB artifacts into memory.