covidcast acquisition rejects valid nan-coded rows

[krivard]

We're getting acquisition failures that may be the primary cause of the dramatic drop in added CHNG rows.

Sample error:

{
  "detail": [
    "Pandas(geo_id='17000', val=nan, se=nan, sample_size=nan, missing_val='4.0', missing_se='4.0', missing_sample_size='4.0')",
    "missing_val"
  ],
  "file": "/common/covidcast/receiving/chng/20210926_county_smoothed_outpatient_covid.csv",
  "event": "invalid value for row",
  "logger": "load_csv",
  "level": "warning",
  "timestamp": "2021-10-04T03:09:38.804771Z"
}

Looking at our tests, it appears that we always instantiate data frames as python objects and not by parsing string data. pandas may be inferring the wrong data type for the missingness columns (here looks like it thinks it's a string), causing the error.

[dshemetov]

It seems to be casting the Nancode to a float and then converting to a string when opening the file, which is an extremely "wat" moment. Fixing now.

[krivard]

Fixed in #733