`split_batch_for_grpc_response` over-splits batches after IPC deserialization

**Describe the bug**

`split_batch_for_grpc_response` uses [this](https://github.com/apache/arrow-rs/blob/54d81919206d667d36ba4b5f76ff8e7ed9d341dd/arrow-flight/src/encode.rs#L621) to estimate batch size for splitting. This calls buffer.capacity(), which reports the full backing allocation size, not the actual data size.
After IPC deserialization, the IPC reader creates all buffers by slicing one shared `Buffer` via [this](https://github.com/apache/arrow-rs/blob/54d81919206d667d36ba4b5f76ff8e7ed9d341dd/arrow-buffer/src/buffer/immutable.rs#L298-L312), which clones the Arc. So every buffer's capacity() reports the entire message body size rather than its own slice.

**To Reproduce**

`FlightDataEncoderBuilder` encode -> `FlightRecordBatchStream` decode -> `FlightDataEncoderBuilder` re-encode produces more FlightData than just `FlightDataEncoderBuilder` encode

**Expected behavior**

**Additional context**

#5352