feat(client): stream ParallelGet in-order to an io.Writer

[worstell]

ParallelGet — and the DownloadGitSnapshot helper that wraps it — previously wrote chunks to an io.WriterAt, which requires a seekable destination (e.g. a temp file) and so prevents a consumer from overlapping the download with processing. They now fetch chunks in parallel but emit in-order bytes to a plain io.Writer via a bounded reorder buffer, letting a streaming consumer (e.g. a decompress/extract pipeline) run concurrently with the download.

A concurrency-sized window caps fetched-but-unwritten chunks, and the reorder buffer is a ring of that many slots, bounding peak memory to O(concurrency * chunkSize) regardless of object size or consumer speed. A chunk whose body length differs from its requested range (short, or overlong from a backend that ignored the range) is rejected rather than spliced or truncated. Revision safety (ETag pinning via If-Range), empty-object handling, range-ignore degrade, and the concurrency == 1 shortcut are unchanged.

The io.WriterAt variant is removed: no consumer benefited from scatter-writes, and the only use — download-to-temp-file then extract — is slower than streaming because it gives up download/extract overlap. *os.File satisfies io.Writer, so the CLI caller is unaffected.

Tests cover in-order reassembly, out-of-order completion, the single-worker/empty-object/range-ignore fallbacks, ETag-mismatch and overlong-chunk rejection, and error propagation, all under -race.