[WIP] Adds partial support for the unofficial bash strict mode

[nkakouros]

• I have reviewed the contributor guidelines.
• I have reviewed the code of conduct.
• Per GitHub's Terms of Service, I am aware that I license my
  contribution under the same terms as this project's license, and
  that I have the right to license my contribution under those terms.

cc: @mbland

I run my scripts with set -euo pipefail at the very top. I run into some unbound variables when doing so. Maybe I'll run into more as I move into using this framework. I think it is good to have default values/explicit null to make assumptions in the code clearer.

On the part where I changed := to :-, I did so to avoid running the same value assignment twice.

This PR is regardless of #220, but it could prove handy if #220 starts getting implemented. This PR is meant to support those users that do want the strict mode in their scripts.

[coveralls]

Coverage decreased (-21.8%) to 73.216% when pulling 2f126ca on tterranigma:unbound into b8d7140 on mbland:master.

[nkakouros]

@go.search_plugins returns 1 some times. Based on the fact that without strict mode, the script executes without errors, I guess that when it returns 1, it is not an actual error. It is probably meant to help when @go.search_plugin is used in an if statement.

If the above is correct, @go.seach_plugin is used in an if statement only once in the whole project, while it is used 8 times as here. Should the function be redesigned or, if the above 'fix' is sufficient, should the same 'fix' be applied in all other 7 case?

[coveralls]

Coverage decreased (-21.9%) to 73.14% when pulling cca38a8 on tterranigma:unbound into b8d7140 on mbland:master.

[coveralls]

Coverage decreased (-22.7%) to 72.306% when pulling c1423d7 on tterranigma:unbound into b8d7140 on mbland:master.

[coveralls]

Coverage decreased (-22.7%) to 72.306% when pulling 6bbf0c9 on tterranigma:unbound into b8d7140 on mbland:master.

[coveralls]

Coverage decreased (-21.9%) to 73.152% when pulling 8201541 on tterranigma:unbound into b8d7140 on mbland:master.

[coveralls]

Coverage decreased (-22.4%) to 72.673% when pulling 7e92dd5 on tterranigma:unbound into b8d7140 on mbland:master.

[coveralls]

Coverage decreased (-22.4%) to 72.663% when pulling d8f6d9e on tterranigma:unbound into b8d7140 on mbland:master.

[coveralls]

Coverage decreased (-22.4%) to 72.663% when pulling d8f6d9e on tterranigma:unbound into b8d7140 on mbland:master.

[nkakouros]

I have a question. When using bash autocomplete, there are around 4,5 commands that return '1' if no candidate for completion is found. When setting set -e, this stops execution at the first such command. This is no problem, as the completion would anyway return nothing in the command line. But when using trap custom_function ERR without set -e, the commands that return 1 cause custom_function to run as many times as there are such commands. This is bad as the user receives unexpected output when instead nothing should be printed in the command line pressing TAB. Should this PR try to 'fix' these commands as well? In my own projects, I take care of the situation in my custom_function by searching for libexec/complete in "${BASH_SOURCE[@]}" and doing set -- ERR` if found.

[coveralls]

Coverage decreased (-22.4%) to 72.675% when pulling 9f38958 on tterranigma:unbound into b8d7140 on mbland:master.

[coveralls]

Coverage decreased (-22.4%) to 72.675% when pulling 0899bfe on tterranigma:unbound into b8d7140 on mbland:master.

[coveralls]

Coverage decreased (-22.3%) to 72.725% when pulling 91dc1c2 on tterranigma:unbound into b8d7140 on mbland:master.

[nkakouros]

Here, it would make sense to me to actually reduce num_completions by one as the array indeed has one less item. But if I do this, then ./go test modules/arg-completion fails on the last test. Why is that?