caching common tokens

[OmarTawfik]

Our scanner copies the string contents of every token when creating it. But many tokens share the same source string:

• Whitespaces.
• Newlines.
• Common keywords like public, contract, function.

We should calculate the most frequent tokens (using sanctuary datasets), and define them as static/const values in source, returning static references when possible, to eliminate allocations:

• We can try using phf::Set::get_key in Token::new() to fetch the static refs and store them instead of cloning the original input, similar to smol_str::SmolStr::new_static.
• We can also try making the whole Token::new() as const, storing the entire object instead of just the string in the phf map.
• We can also try caching the Rc<Token>, so that all of them share the same instance/pointer.

Once we have #808, we can evaluate the impact of these optimizations.

[AntonyBlakey]

I think on-demand interning will be enough, without pre-computation