### What happened?

When using LiteLLM Proxy's Anthropic-compatible `/v1/messages` endpoint with a Bedrock Claude model through `bedrock/converse`, streaming stalls during large `tool_use` calls.

If the model emits a large tool call argument payload, for example a `Write` or `Edit` tool call with a large `file_text` argument, LiteLLM sends the initial SSE events normally, then stops sending bytes for tens of seconds or longer. After the tool argument generation finishes, LiteLLM flushes many `input_json_delta` chunks all at once.

This causes clients or intermediaries with idle timeouts to close the stream before the delayed chunks arrive.

### Reproduction

1. Configure LiteLLM Proxy with a Bedrock Claude model using the Bedrock Converse path. 2. Send a streaming request to `/v1/messages`. 3. Use a prompt/tool setup that causes the model to emit one large `tool_use` block, such as writing or editing a large file. 4. Observe raw SSE timing from the client side or proxy side.

### Observed behavior

The stream starts normally:

```text t=0.00s event: message_start t=0.05s event: content_block_start (text) t=0.12s event: content_block_delta (text) t=0.18s event: content_block_st…