Allow custom shell code to process Swift custom completion candidates

Allow custom shell code to process Swift custom shell completion candidates.

Currently, the completion candidates returned from Swift custom shell completion code to the requesting shell are only processed in a single way per shell. Sometimes, you need more complex shell-side scripting to parse more complex completion concepts.

e.g., I want to match the completing shell word against the Swift `String` `name` property of certain structs, but, once a completion candidate has been selected, I want the text that is inserted into the command line to be the integer `id` property from the same struct. I cannot do this using the current embedded call to the zsh `_describe` function.

The solution is to add a `shellScript: String = ""` parameter to `CompletionKind.custom(…)`.

When the shell completion script is generated, the value of `shellScript` will be escaped for a single-quoted shell string, and then be inserted into a single quoted string. If that string is empty, existing completion candidate processing will be used; otherwise, the given shell script will be run via `eval` to process the custom completion candidates when completion candidates are requested by the user.

The custom shell script will be able to access the completion candidates from a shell array, tentatively named `completions` (for zsh, there will be 2 arrays `non_empty_completions` & `empty_completions`). The custom shell script will be responsible for matching, displaying & inserting candidates.

if the `completion` closure is not provided, it will default to `nil`, in which case the generated completion script will eval the `shellScript` without calling Swift for completions (i.e. it won't call the Swift binary with a `---completions` option). In this case, all completion candidate arrays available to `shellScript` will be empty.

The `shellScript` can be completely different for different shells; Swift code would need to switch on `CompletionShell.requesting` for different shells & on `CompletionShell.requestingVersion` for different versions of a shell.

In the future, if there are common complex reusable completion candidate processing patterns, functions returning a `String` of shell code could be added to SAP so that users could use them instead of the default processing without needing to write the scripts themselves, as they would probably be complicated, and producing similar functionality across 3 shells would be even more difficult.

Start by locating CompletionKind.custom(…) and the generated shell completion logic for the supported shells. Trace how custom completion candidates are currently passed and processed, then define the shellScript and nil-completion behavior described in the issue. Done means custom scripts can access candidate arrays, while empty scripts retain existing processing across shell versions.

Written by the indexing model from the issue text.