Improve error messages when reference resolution fails

[simu]

To improve the error messages for reference resolution failure, we add a Vec<String> to track the current parameter key for which we're resolving references into the ResolveState struct.

We also extract the error rendering in Token::resolve() into methods on ResolveState, since we mostly need data from ResolveState to render the error messages.

The new error messages all print the parameter key for which we're resolving references. If applicable, the error messages also contain the full reference, and the particular key for which resolution failed.

To reduce repetition for the error handling, we also refactor Token::resolve() to always either interpolate or copy the Value which is the base for the key lookup. While this introduces a bunch of unnecessary calls to clone() if we're not doing a lookup into a Value::String or Value::ValueList, the performance impact is negligible, and the resulting code is much less convoluted.

Resolves #69

Checklist

• The PR has a meaningful title. The title will be used to auto generate the changelog
• PR contains a single logical change (to build a better changelog).
• Update the documentation.
• Update tests.
• Categorize the PR by setting a good title and adding one of the labels:
  bug, enhancement, documentation, change, breaking, dependency, internal
  as they show up in the changelog
• Link this PR to related PRs or issues.

[simu]

Note that the PR has a slight performance regression (the benchmarks in #47 report a performance regression of 2-4%).

As far as I'm able to tell, this is caused mostly by the additional state that we're tracking:

1. The ResolveState struct is larger, which makes cloning it during reference resolution more expensive
2. We introduce a bunch of extra allocations because we track the current key in a Vec<String> -- note that tracking the current key in a String isn't faster.

[simu]

Follow-up regarding performance: I tested a bit more and have found that pre-allocating the variable-sized ResolveState fields (seen_paths: HashSet::with_capacity() and current_keys: Vec::with_capacity()) doesn't seem to make a significant difference in runtime.