FSQ implementations' bug

[scenarios]

Your implementation on bound function might be incorrect.
tan() is not the inverse of tanh().

[lucidrains]

ah yes, i see it

the reason is bc it was transcribed from the pseudocode in the paper

@fab-jul do you have any thoughts?

[lucidrains]

#95 will merge if author confirms this is an error

[fab-jul]

Hi yes this is a mistake, should be arctanh.

Note that it barely matters as it's called with a value around 0, where both are similar.

Also note that the whole bound function is really ad-hoc and should be revisited. for example, maybe it should be symmetric around 0 also for even numbers. If anyone has some insights let me know.

[lucidrains]

@fab-jul yea i figured as i hear from a lot of people having success with FSQ! let me get this merged in just for correctness sake

@scenarios thank you for spotting this