refactor(_command_offset): save Bash's $2 passed to completion functions

[akinomyoga]

Fixes #790

This is a fix alternative to #791. Instead of imitating the arguments passed by Bash, we may save the original arguments in _init_completion (renamed to _comp_initialize). However, this involves not only the change of _init_completion itself but also the interface change of functions that call _init_completion internally---such as _filedir_xspec, _longopt, _minimal, _known_hosts, etc. Also, this doesn't work when _comp_command_offset is used without calling _init_completion (some cases are found in external completions). We need to request external completion writers carefully update their code.

[scop]

Even though this PR is huge and there is an interface change involved, I feel the code here is much easier to understand than in #791; from that point of view I have a better feeling about this one.

If I understand correctly, the API change is such that it will mostly degrade gracefully, i.e. if the original arguments are not available, we'll basically do what we used to do before this change? Or will this flat out break some scenarios if not adjusted by affected completions?

External completion authors will likely be using _init_completion for a long time still. Luckily the current implementation (before this change) should be fine if it receives positional arguments, so I suppose externals can switch basically to

    local cur prev words cword split comp_args
    _init_completion -- "$@" || return

...in order to support earlier bash completion versions and to take advantage of the new stuff.

[akinomyoga]

Even though this PR is huge and there is an interface change involved, I feel the code here is much easier to understand than in #791; from that point of view I have a better feeling about this one.

OK! Then, let us consider this direction!

If I understand correctly, the API change is such that it will mostly degrade gracefully, i.e. if the original arguments are not available, we'll basically do what we used to do before this change?

Yes, it is basically right. Actually, it slightly improves the original behavior. Originally, we passed the entire word ${COMP_WORDS[COMP_CWORD]} as the second argument, but now we are passing the part of the word after the current cursor position, i.e., $cur, as Bash does for its built-in programmable completion.

External completion authors will likely be using _init_completion for a long time still. Luckily the current implementation (before this change) should be fine if it receives positional arguments, so I suppose externals can switch basically to

    local cur prev words cword split comp_args
    _init_completion -- "$@" || return

...in order to support earlier bash completion versions and to take advantage of the new stuff.

Ah, this is a good point. I hadn't noticed that point by myself. Thank you.

[scop]

Ok, let's go with this. Thanks!