fix(downloader): keep the chunk in memory and retry the write instead of re-downloading

[lukyrys]

When a chunk write fails because the disk is full or the target directory is not writable, the downloader now keeps the already-verified chunk in memory and retries the same bytes instead of dropping them and downloading the chunk again.

• the retry runs on the existing WRITE_RETRY_DELAY interval up to MAX_WRITE_RETRIES; all other writes for that LISH stay paused meanwhile
• a peer that hits the error while another peer already owns the pause waits for it and then retries its own retained buffer, so concurrent peers cannot burn the shared retry limit against each other
• if the file disappears during a retry (ENOENT), the chunk goes back to the queue and the existing missing-file recovery takes over
• the reported error code follows the live cause, so an EACCES that later turns into ENOSPC is reported correctly
• no frontend changes