Inverted split_off logic in IterSource::seek corrupts resumed unbuffered upload

Unconfirmed. This is the result of a search with Claude, may be a false positive.

`IterSource` backs `Payload::from(Vec)` during unbuffered and streaming uploads. On partial resume, `seek(offset)` is intended to keep the unpersisted suffix `[o, n)` of the buffer element containing the offset.

However, in `src/storage/src/storage/streaming_source.rs`:

https://github.com/googleapis/google-cloud-rust/blob/f02b42c2ac5720334cd2b5335824bd4166264527/src/storage/src/storage/streaming_source.rs#L313-L316

The code executes `b.clone().split_off(n - o)`, which retains the last `o` bytes instead of the tail `n - o` bytes. This leads to the same data corruption on resumed unbuffered uploads as seen in the buffered upload path.

Start in src/storage/src/storage/streaming_source.rs at IterSource::seek, especially lines 313-316, and trace how the offset and buffer lengths determine the unpersisted suffix. Confirm the behavior against resumed unbuffered uploads; done means the suffix after the resume offset is preserved without corrupting the upload.

Written by the indexing model from the issue text.