Adds support for GPT-5 series of models in `arcade docs`.
Improves error handling when the LLM does not generate a valid JSON for
a given tool sample inputs. Instead of raising an exception, the CLI
uses an empty input, moves on to the next tool, and prints a warning
message asking the user the fill in the input sample manually in
Javascript and Python files.
This PR also moves the Enumerations from a separate `reference.mdx` file
to the main toolkit file, as requested by @EricGustin to simplify the
docs structure.