[FEA] Allow Configuring Default Bit Width Used for Json and Csv Readers

[isVoid]

As part of #10558 , workflows can have benefits of saving memory if a default bitwidth can be configured for the columns read via cuio. Today, csv and json data format does not encode the data type in file and is defaulted to 64bit if no data type is specified. It would be nice to provide a reader option to default it to 32bit.

It would also be nice to have a separate option for integer columns and floating point columns, because the precision requirement for these two column types are often different.

An example use in cudf python is done via a global config:

cudf.set_config("default_io_int_bitwidth", 32)
df = cudf.read_csv(file) # all integer columns are 32 bit

An alternative to this approach today is to read a small portion of the file, determine the column kind and limit the precision in the second pass:

dtypes = cudf.read_csv(file, nrows=1).dtypes
dtypes = [np.dtypes(f"{dtype.str[:2]}{32}")]
df = cudf.read_csv(file, dtype=dtypes)

This is tedious setup and requires an extra kernel launch. It would be nice to avoid that.