Improve memory management

[z-silver]

Right now, we don't have a proper strategy for allocation during parses. Most likely we are leaking memory everywhere. We can do better.

After looking into how std.json is implemented, I think we can largely borrow their memory management strategy: instead of direcly returning the user-requested type T, we could return something like std.json.Parsed(T) which wraps the T together with an arena allocator which is used for parsing.