refactor(_command_offset): refactor the adjustment of COMP_WORDS

[akinomyoga]

Extracted from #790 #791. This is also actually independent.

• d2c0b4e simplify logic: The original code tries to manually perform a string manipulation that can be achieved by a single parameter expansion. Also, it tries to manually perform the array compaction that can be achieved by a single parameter expansion. The behaviors of the new codes slightly differ from the original ones, but I believe the behaviors of the new codes are better:
  • In the original code, COMP_LINE becomes empty when ${COMP_WORDS[i]} is not contained in $COMP_LINE, though I don't think that happens as far as these variables are modified by external codes. In the new code, such a word ${COMP_WORDS[i]} is just ignored.
  • In the original code, the words in COMP_WORDS after COMP_CWORD is not modified, but just the word at COMP_CWORD and the words preceding it are shifted. For example, COMP_WORDS=([0]=sudo [1]=A [2]=B [3]=C [4]=D) with COMP_CWORD=2 becomes COMP_WORDS=([0]=A [1]=B [3]=C [4]=D) instead of COMP_WORDS=([0]=A [1]=B [2]=C [3]=D). In the new code, all the words are shifted.
• 995db59 localize changes to COMP_*: The original implementation of _comp_offset breaks the state of COMP_* after the completion of _comp_offset. Currently, it doesn't cause a problem because all the completion functions in bash-completion that use _comp_offset seem to return from the function immediately after the call of _comp_offset. But I still think we should localize the changes to COMP_* within _command_offset and functions called from it.
  • This change might possibly break existing external completion functions if there are completion functions that use the modified values of COMP_* after calling _comp_offset.
  • The test seems to ignore also other variables cur, words, and cword, and they do not seem to be modified by the current implementation of _command_offset.

The description is the same as was in #791.

[scop]

Thanks!