parse richInspectVariables result with ast.literal_eval

[Keshava-kesh]

When the kernel is stopped on a breakpoint, richInspectVariables asks debugpy to evaluate display_formatter.format on the variable and gets the answer back as a string, the repr of a (data, metadata) tuple. That string is then handed to eval to turn it back into two dicts. The bundle values are whatever the object's formatters returned, so any value whose repr is not a plain literal gets executed as an expression a second time, in the control thread, instead of being read as data. A repr_mimebundle that returns an object with a repr of len('abc') comes back from the handler as 3. I ran into it reading through the debug handlers. Parsing the result with ast.literal_eval keeps the normal case identical and leaves the representation empty when the string is not a literal, the same way the non-breakpoint branch does when user_expressions fails. The change stays inside richInspectVariables, with a regression test next to the existing breakpoint one.