[common] Fix snapshot and tag reads failing when the file is deleted concurrently

[sundapeng]

Purpose

Paimon treats a snapshot, tag, changelog or schema file as missing only when reading it throws FileNotFoundException. Some FileIOs, such as OSSFileIO and Jindo, throw a plain IOException instead when the file is deleted while it is being read. A writer traversing snapshots in latestSnapshotOfUser then fails its checkpoint when it reaches a snapshot being expired, and TagManager#get fails instead of returning empty for a tag being deleted.

When reading fails with another IOException, FileIO#readFileUtf8 now throws FileNotFoundException if the file is gone, and keeps the original error otherwise.

Tests

FileIOTest: 2 new cases, 7 tests passed. SnapshotManagerTest: 1 new case, 43 tests passed.

[LsomeYeah]

Thanks for the contribution! I wonder whether it would be more appropriate to normalize this kind of missing-file exception at the FileIO layer？

SnapshotManager may not be the only caller that relies on FileNotFoundException for special handling. For example, incremental scans also distinguish a missing tag by catching FileNotFoundException before falling back to snapshot IDs. Perhaps handling this in FileIO could provide consistent behavior for snapshots, tags, and other metadata files.

[sundapeng]

Thanks, that makes sense. Tag, Changelog and SchemaManager have the same gap, and all of them read through FileIO#readFileUtf8, so I moved the check there: when a read fails with an IOException other than FileNotFoundException and the file no longer exists, it throws FileNotFoundException. readOverwrittenFileUtf8 already does a similar exists() check for hint files.

I kept it in the default method instead of the OSS and Jindo FileIOs, because those can't tell a missing file from other failures either: hadoop-aliyun drops the OSS error code and fails with Null IO stream.