Use Pandas' `read_json` in the JSON builder

Instead of PyArrow's `read_json`, we should use `pd.read_json` in the JSON builder for consistency with the CSV and SQL builders (e.g., to address https://github.com/huggingface/datasets/issues/5725).

In Pandas2.0, to get the same performance, we can set the `engine` to "pyarrow". The issue is that Colab still doesn't install Pandas 2.0 by default, so I think it's best to wait for this to be resolved on their side to avoid downgrading decoding performance in scenarios when Pandas 2.0 is not installed.