Refactor the RPC MetadataResponse to introduce invalid partition set

### Search before asking

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

### Description

Currently, when a `MetadataRequest` is sent, if the requested tableBucket contains partitions that do not exist, fluss server will directly throws an RPC-level exception: `PartitionNotExistException`. This makes it impossible for the client to handle the error gracefully. For example, the `LogFetcher` would crash immediately, affecting the reading of other normal tableBuckets.

To address this, it is proposed to return a set of invalid partitions (i.e., all non-existent partitions) instead of throwing an exception. This allows the `LogFetcher` to report these invalid partitions to the user, who can then customize the next fetch logic (e.g., skip or retry specific partitions).

### Willingness to contribute

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

Trace the MetadataRequest handling and the LogFetcher behavior described in the issue. Determine how the MetadataResponse should represent non-existent partitions without an RPC-level exception, then verify that invalid partitions can be reported while other normal tableBuckets continue to be read.

Written by the indexing model from the issue text.