suggest using exec foo for top-level commands rather than foo for consistency

[mvdan]

At https://github.com/cue-lang/cue we use test scripts for our integration tests, as well as for users providing reproducers for bugs.

There's one unfortunate inconsistency: the former use cue, and the latter use exec cue.

I initially thought that we should use cue everywhere, by adding a new flag to cmd/testscript like -x cue, so that a command "passes through" to executing a program from $PATH. Then, we would use cue everywhere rather than exec cue. However, then bug report reproducers aren't truly standalone, as one needs to remember to use the flag.

@rogpeppe suggests that we should use exec cue everywhere instead. This already works in today's testscript.RunMain, as top-level commands are put in a temporary directory and appended to $PATH. We simply prefer cue rather than exec cue since we're used to how it was before, where exec cue might fail to find the program if one hadn't run go install first, as we wouldn't fix up $PATH.

I think Roger's idea is better. It's slightly more verbose, but:

1. It's consistent everywhere - tests and standalone reproducers.
2. It keeps exec explicit, which is a good thing given that it's special. It uses stdin, sets stdout and stderr, runs as a separate process, can be run in the background, etc.

[mvdan]

Also: we should add a boolean flag like RequireExplicitExec so that users of RunMain can consistently use exec cue and not make the mistake of using cue inconsistently. RequireExplicitExec would make cue error, telling the user to use exec cue instead.

[myitcv]

You've won me over 😄