Allow maximum deviation while supporting all f32 values for Swing

[skejserjensen]

Currently maximum_allowed_deviation() returns a value that is lower than the error bound because inaccurate floating point math seemingly can cause deviations that are higher then the error bound when converting the first and last value to slope and intersect for Swing (it looks like it can still occur in rare cases). Since this may reduce the amount of compression that can be done, and even worse, may allow values to be stored outside the error bound, the min/max values should be converted to slope/intercept in a way that allows maximum_allowed_deviation() to always return the maximum deviation. This may require min/max to be stored as f64 instead of value as value seems to be too inaccurate to recreate the original slope/intercept.