.NET Core 3.0 is unsupported now and targeting it actively throws a warning on the .NET 5 SDK.

For future commit to add Half support.

Float serialization used the same serialization code as normal integers, which writes less bytes if the integer is of sufficiently small size. The problem however is that a float is basically *never* so small due to how the bits are laid out and this almost always resulted in a wasted extra byte vs just writing the 4/8 bytes of the float directly.

("basically never" is everything except 0/subnormals, which I'd say is not worth adding an extra byte to almost all other floats over. I even tested all possible floats just to make sure.)

The new code is also faster on Core because it writes the 4/8 bytes at once instead of individual WriteByte/ReadByte() calls (see benchmarks in PR).