InferenceClient: handle error response in when streaming tokens from TGI

[Wauplin]

In InferenceClient.text_generation when using a TGI endpoint with stream=True, it is possible that a serialized error is returned instead of a token (e.g.: out of memory error). Once an error is received, we should parse it correctly + raise a correct Python error. At the moment, a Pydantic error is raised because it cannot be validated as a TextGenerationStreamResponse schema.
Once an error is received, we can safely assume that no more tokens will be received in the stream.

The rust schema is:

#[derive(Serialize, ToSchema)]
pub(crate) struct ErrorResponse {
    pub error: String,
    pub error_type: String,
}

The logic to parse a stream of tokens from TGI can be found here:

huggingface_hub/src/huggingface_hub/inference/_common.py

Line 263 in 89cc691

def _stream_text_generation_response(