fix(_comp_command_offset): complete command name when called for it

[ybai08]

Bash 4.3..5.3 may call the completion function of a meta-command for the command name itself, with COMP_CWORD=0 (as found in #1383, where it was identified as a Bash bug). For A= ./d[TAB]o, the function for do (_comp_command) is called with COMP_LINE=./do, COMP_POINT=3, COMP_WORDS=(./do). _comp_command passes offset 1 to _comp_command_offset, which points past the current word:

• A= ./d[TAB]o: COMP_POINT becomes -1, and _comp_get_words prints bash: COMP_POINT: substring expression < 0 (COMP_POINT: substring expression < 0 when completing A= ./do #1383).
• A= ./do[TAB] ;: COMP_WORDS becomes empty and nothing is completed; with nounset it prints bash: COMP_WORDS[0]: unbound variable (COMP_WORDS: bad array subscript #768; the COMP_WORDS[-1] line quoted there has since changed, but the empty COMP_WORDS remains).

This clamps the offset to the current word, so the word is completed as a command name through the existing COMP_CWORD == 0 branch (_comp_compgen_commands). That matches what Bash does with no bash-completion loaded: in both cases ./dofile (an executable) is completed.

Callers other than _comp_command/_comp_root_command go through _comp_initialize, which already returns early for cword <= 0, so they do not reach this path.

Closes #1383
Closes #768

Testing

New test_cword0 cases in test/t/unit/test_unit_command_offset.py emulate the Bash call (COMP_CWORD=0, COMP_WORDS=(./com)) for the two cursor positions above and expect ./completer from the _command_offset fixture dir.

Run on macOS with Bash 5.3.0 built from the GNU release tarball (BASH_COMPLETION_TEST_BASH):

• Without the bash_completion change, both new cases fail: Unexpected output: [bash: COMP_POINT: substring expression < 0 and Unexpected output: [bash: COMP_WORDS[0]: unbound variable. With it: t/unit/test_unit_command_offset.py 17 passed.
• Interactive check with pexpect (Bash 5.3, --norc, only bash_completion sourced): A= ./d[TAB]o, A= ./do[TAB] ;, var= ./do[TAB] ; and A= ./do; x with the cursor after ./do now complete ./dofile the same way as Bash without bash-completion; A= ./do[TAB] and nice ./do[TAB] are unchanged.
• t/unit and 691 of the 730 t/test_*.py files (ssh/scp/rsync, network and database client tests excluded) with -n 8: the same pre-existing macOS failures as on main (test_aclocal, test_tar::test_25, test_vipw, intermittently test_ps/test_cpio, and environment-diff errors from bash: child setpgid ...: Operation not permitted), nothing new.
• shellcheck -f gcc bash_completion (0.11.0) and shfmt -d bash_completion (3.14.1) are clean; ruff check / ruff format --check (0.15.20) clean on the test file.

Not run: Bash versions other than 5.3 (the CI matrix covers those), test/runLint (needs GNU grep -P), the excluded test files, and prek/mypy.

Side note, unrelated to this change: test_cmd_quoted in the same file fails when run on its own on main (pytest t/unit/test_unit_command_offset.py::TestUnitCommandOffset::test_cmd_quoted), because it relies on test_2[cmd2] having replaced the cmd2 compspec first in the shared class-scoped shell. With -n it can fail depending on how tests are distributed across workers.