fix: don't truncate compressed packets that outgrow the output reserve

[luizribeiro]

The existing code assumes a compressed packet is always smaller than the uncompressed one: Compress::compress_into sizes zlib's output buffer at input.len() + 10 bytes and works from there.

That isn't true when the data is already compressed. For pack files, encrypted blobs or media, zlib has nothing left to squeeze out, so it wraps the bytes in a stored block and hands back slightly more than it was given. Past 32 KiB the stored-block overhead is more than the 10 bytes of slack, so the buffer is too small.

Overflowing is supposed to be recoverable: the loop grows the buffer and calls zlib again. But it only grows on flate2::Status::BufError, and zlib returns BufError only when it is completely stuck. Filling the buffer while still making progress comes back as Status::Ok, which the loop reads as "done", so it sends the packet with the tail still inside zlib. The packet arrives shorter than its header claims, and the bytes left behind are prepended to the next one, so the stream stays broken from there on. OpenSSH reports it as channel 0: get data: incomplete message: a git clone against a russh server dies as soon as pack data starts flowing, while ordinary traffic is unaffected because it does compress.

The fix stops asking zlib whether it is stuck and asks whether it left any room. Room to spare means it emitted everything; a buffer filled to the brim means there may be more, so grow and call again. compress now delegates to compress_into so the two copies of the loop can't drift apart again.

The new tests round-trip packets that outgrow the reserve, including through compress_into at a non-zero start_len, the shape the packet writer uses. The existing tests only cover compressible payloads under 4096 bytes, which always fit. The new ones fail without the fix (payload of 32769 bytes came back as 32735 bytes) and pass with it; the full suite passes. Verified end to end too: a stock git clone over a compressed session against a russh-backed git server fails on 0.62.6 with exactly that error and succeeds with this change.