Batch result decoders leave responses open after parsing or transport errors

When a batch-result JSONL stream raises a JSON decoding or transport-read error, its HTTP response stays open. The caller catches the error, but the connection is not released unless they also explicitly close the result decoder.

Reproduced on main `eb21a4352015686c30f5759e8c2f02d70f5371e2`, Python 3.10/Linux, for both sync and async `client.messages.batches.results()`. A minimal reproducer of the decoder used by that endpoint is:

```python
import json
import httpx2
from anthropic._decoders.jsonl import JSONLDecoder

response = httpx2.Response(200, stream=httpx2.ByteStream(b"invalid json\n" + b" " * 64))
decoder = JSONLDecoder(
raw_iterator=response.iter_bytes(chunk_size=64),
line_type=object,
http_response=response,
)
try:
list(decoder)
except json.JSONDecodeError:
pass
print(response.is_closed) # False; expected True after the iterator fails
response.close()
```

The same leak happens if the underlying byte stream raises `httpx2.ReadError`. Successful exhaustion already closes the response through HTTPX. The error paths should release it as well, matching the SDK's SSE stream cleanup.

AI assistance was used for investigation and local verification; no API calls were made to a live provider.