Roundtrip orient for JSON

[Komnomnomnom]

Add a round-trippable json orient (maybe "roundtrip") to provide enough metadata to reconstruct a frame or series with 100% fidelity. As discussed in #9130

[Komnomnomnom]

This could also support multi-level indexes and panels etc.

[cpcloud]

@jreback and i are currently using a schema that looks somewhat like this:

schema = {
    'pandas': {
        'type': 'DataFrame',
        'version': pd.__version__,
        'orient': 'records',
        'date_unit': 'ms',
        'values': df,
        'dtypes': {
            'str': [dtype.str for dtype in dtypes],
            'characters': [dtype.char for dtype in dtypes],
            'kind': [dtype.kind for dtype in dtypes]
        },
        'axes': {
            'columns': {
                'values': df.columns.values,
                'dtype': df.columns.dtype.str,
                'names': df.columns.names
            },
            'index': {
                'values': df.index.values,
                'dtype': df.index.dtype.str,
                'names': df.index.names
            }
        }
    }
}

[jreback]

@Komnomnomnom

the idea is the encompas enough meta data to really make a DataFrame round tripable w/o passing anything, IOW, making it a fully-capable format. Of course its a 'user-made' one, but that is true with lots of formats.

I suspect lots of people have already implemented this (painfully) by creating their own formats. This is basically trying to set a standard, in that, pandas objects become JSON serializable in a well-defined format.

This can be done simply by using pandas.io.json.dumps with this format on the writing. @cpcloud mentioned that it might be necessary to support some sniffing the reader to de-serialize this format (e.g. look for the pandas tag then use the meta-data to set the next part of the serialization options).

The idea of the format is that it is pretty straightforward to extend to Series/Panel/Multi-Index.