RemoteLogDownloader deadlock when met exception.

### Search before asking

- [x] I searched in the [issues](https://github.com/apache/fluss/issues) and found nothing similar.

### Fluss version

0.7.0 (latest release)

### Please describe the bug 🐞

In the original code, the download lock (prefetchSemaphore) is released only in two cases:

1. After a RemoteLogSegment has been successfully read (drained), the lock is released via recycleRemoteLog.

2. When the download of a file fails, the lock is released.

Let us simplify the model: suppose a bucket contains three segment files — A, B, and C — and client.scanner.remote-log.prefetch-num = 1.
1. File A fails to download, so the lock is released. File A is then added back to the end of the queue.
2. File B downloads successfully, but the lock is not immediately released because it hasn't been drained.
Since file A has an earlier offset, it remains at the front of the queue and must be processed first. However, file B holds the prefetch lock, and file A cannot be reattempted until the lock is acquired again. But because B will never be drained (as A blocks its processing), the lock is never released — resulting in a deadlock.
3. file C will not be downloaded, and file A will never be retried. The entire job becomes stuck.

### Solution

When RemoteLogDownloader failed to download a file, no longer release the semaphore but retied to download for several times. If still failed to download, but thrown the exception out of client, let the flink job fails and restarts.

### Are you willing to submit a PR?

- [x] I'm willing to submit a PR!

Start at RemoteLogDownloader and trace how prefetchSemaphore is acquired, released, and how failed files are requeued. Reproduce the A/B/C ordering with prefetch-num set to 1, then verify that repeated download failures eventually surface an exception instead of leaving the downloader deadlocked.

Written by the indexing model from the issue text.