Size optimization for Option

[dragazo]

Closes #47.

As far as I can tell, there was no real need to use MaybeUninit<InlineString> instead of just InlineString directly, due to it being a POD type. All that was required was changing the types of pointer casts we do to directly use &self.data as *const _ rather than self.data.as_ptr() in various places.

Due to no longer using the MaybeUninit union type, we are able to get Option size optimization out of the InlineString. To do this, I changed Marker(u8) to Marker(NonZeroU8) and changed it to use the low 2 bits for storing both the discriminant (as it already did), and also a non-zero bit to guarantee the requirement of NonZeroU8. I've also changed the allocator alignment requirement to 4 instead of 2 to guarantee these bits are free to use; this should not cause issues on 32 or 64 bit systems, which are our targets.

To ensure that a BoxedString write over the memory of the InlineString does not violate the requirements of NonZeroU8 (and/or produce a None value), the ptr field in BoxedString was changed to a new TaggedPtr type that automatically handles setting the low 2 bits appropriately (and removing them when queried).

I feel that additional testing should be done, but it currently passes all existing test cases and a few that I added specifically about size requirements and Some vs None checks for small and large strings.

[dragazo]

I also just took the liberty of solving #33 by using alloc::borrow::Cow, which is available in no_std mode.

[dragazo]

I'll also note that this PR now uses 2 bits for the expanded discriminant, but only 3 states (None, InlineString, and BoxedString). The fourth pattern could be used to implement a feature like storing a &'static str like some other sso libs can do. That's outside the scope of this PR, though.