No alloc experiment

[daxpedda]

This adds support for removing the requirement for alloc from the crate. Following changes were introduced:

• All serialization returns GenericArray<u8, _> instead of Vec<u8> and requires some where bounds, which shouldn't affect the user unless they are using a generic interface themselves.
• NonVerifiableClient and VerifiableClient don't hold data anymore, that way the user has to store the data and we don't have to bind our items by lifetimes. This also makes serialization without allocation possible, because we have a known size now. Alternatively we could either introduce a lifetimes to these items, or introduce a generic that can be AsRef<u8> to own the data without allocation. Both would make serialization without alloc support impossible or difficult.
• batch_evaluate is now only supported with alloc, because I have no clue how to support it without allocating memory to save the generated EvaluationElements.
• batch_finalize now returns an Iterator instead of a Vec.

So the missing part to make it completely support building without alloc would be to split the batch_evaluate functions into two parts so output can be collected by the user without allocation. This is already done internally, so we can just make the interface public. We can still leave the original batch_evaluate as a convenience function for users that have support for alloc. What do you think @kevinlewi?

This builds on top of #34.

[daxpedda]

Cherry-picked some commits that don't really belong here to #34.

[daxpedda]

Re-based and ready to review, the question remaining in the OP can be handled in a different PR too.

[daxpedda]

Re-based. CI is fine, failed on downloading Rust.

[daxpedda]

Merged by #44.