Load RTTMs with keep_default_na=False to avoid nan speaker label

[cifkao]

Currently, load_rttm uses pd.read_csv() with keep_default_na=True. This results in speaker labels such as "<NA>", "nan", "NA", "N/A", and "NULL" all being parsed as the float nan value.

It appears that the only value that could legitimately represent a NaN/missing value in an RTTM file is <NA>, but it seems that it would be unexpected for most (meaningful) columns, and it definitely seems incorrect to read it as a float for string columns (uri, speaker).

The fix I'm proposing is therefore to pass keep_default_na=False and na_values=["<NA>"], and convert nan values to None for uri and speaker.