Add values_dtype backend option to load values at full precision

[juntyr]

Supersedes #395

#395 (comment)

Hi @juntyr and @dmey, thanks for this change! However, I think this is something we would prefer to be an option since it will double the memory requirement (and we do have GRIB fields with over 1 billion values).

Potential API:

ds = xr.open_dataset('data.grib', engine='cfgrib', backend_kwargs={'values_dtype': 'float64'})

Do you think you could make this modification to the PR, or would you rather leave it as a request issue?