streams: add a bulk-read operation to StreamIO to unblock perf download throughput

[adust09]

Problem

StreamIO only exposes streamReadByte, so every consumer reads one byte at a time (with an IORef-buffered chunk underneath in the yamux adapter, plus a BS.tail allocation per byte). For message protocols this is fine, but the perf test app (#130, PR #275) drains multi-megabyte payloads through this path:

• upload hs→nim measured 0.49–0.80 Gbps (write path is chunked, fast)
• download nim→hs rounds to 0.00 Gbps (< 5 Mbps) — entirely bounded by the per-byte read path

The unified-testing perf harness defaults to UPLOAD_BYTES/DOWNLOAD_BYTES = 1 GiB, which is impractical at this read speed, so this blocks the upstream perf/images.yaml submission.

Proposal

Add a chunk-level read to StreamIO (e.g. streamReadChunk :: IO ByteString, returning whatever is buffered, EOF via exception as today) and implement it in:

• the yamux adapter (yamuxStreamToStreamIO — hand back the buffered chunk instead of a byte)
• the Noise session wrapper (noiseSessionToStreamIO — a decrypted frame is already a chunk)
• the in-memory test pairs (mkMemoryStreamPair, mkEofStreamPair)

then switch readExactBounded, perf's drainUntilEof/discardExactly, and other bulk readers onto it. streamReadByte can remain as a default derived from the chunk read for the message protocols.

Acceptance

• perf self-test download throughput within the same order of magnitude as upload
• both hs↔nim perf cross directions still pass