Don't hardcode fill_value

[ocefpaf]

Instead of using a float value and relying on old numpy's behavior we will use netCDF4's default fill values:

{'S1': '\x00',
 'i1': -127,
 'u1': 255,
 'i2': -32767,
 'u2': 65535,
 'i4': -2147483647,
 'u4': 4294967295,
 'i8': -9223372036854775806,
 'u8': 18446744073709551614,
 'f4': 9.969209968386869e+36,
 'f8': 9.969209968386869e+36,
 'c8': 9.969209968386869e+36,
 'c16': 9.969209968386869e+36}

This will make pocean-core really numpy 2.0 compliant, even though we didn't catch this one in the tests, and make the data saved smaller and "more consistent" in same cases as the original dtype will be using for the fill value.

[ocefpaf]

I don't see a reason to have a default value, maybe faster writes? If that is true we should check the dtype against netCDF4's defaults. IMO, this solution is "good enough" and will fix some errors we are seeing in ioos-qc when dealing with integers b/c numpy 2.0 won't upcast but will overflow instead.