Add `new` builtin command to generate scripts, fix latent bugs

[mbland]

Closes #142.

The process of testing revealed so much more than just early bugs in the new command script itself. See the commit messages for the gory details, but the highlights are:

• Previously the version check at the very beginning of go-core.bash would report that Bash 4.0 and 4.1 were unsupported. This has been fixed.
• I uncovered another Bash array initialization bug, this time related to brace expansion, that caused several test cases from tests/new.bats to fail on Bash versions prior to 4.1. As with the other array bug from commits b421c7382fc1dafb4d865d2357276168eac30744 and c6bf1cf46c7816c969a0c5d45a4badeb50963f95, the fix was to declare the array on one line and initialize on the next, rather than declaring and initializing it on the same line.
• Debugging test failures on Ubuntu and Arch Linux led me to discover a silent failure in @go.test_compgen whereby the underlying compgen would, in some cases, happily return a successful status even when no completions were generated. @go.test_compgen now fails loudly in this case, as you practically never want it to return empty results (otherwise just use the empty string as test data directly).
• The big one: I discovered a subtle bug in several tests/new.bats test cases only because of a difference in array expansion between Bash versions up to and including 4.2.31 and Bash versions from 4.2.32 on—only when a single-item array is expanded such that it results in the empty string. See the commit message for 5abb59b686ef13894f6979a11931ce007a58e850 for all the juicy details.
• Finally, running tests/new.bats on Windows using the MSYS2 Bash from Git for Windows revealed a bug in stub_program_in_path from lib/bats/helpers whereby stubbing a system program like chmod may result in the program stub itself getting executed as part of create_bats_test_script. The fix was easy enough, by calling create_bats_test_script before setting PATH.

[coveralls]

Coverage decreased (-0.8%) to 94.477% when pulling f530a603252466db23ed659784b3fccd07130774 on generator into bd8994d94d715e8ed021d4fcb4a764aee0d5c55a on master.

[mbland]

Note that the "decrease" in coverage is because kcov isn't recognizing any of the array assignments in libexec/new as executed code. Coverage of that file is actually 100%, and coverage of the remaining files remains unchanged. Merging.