Adjust coverage of Travis builds

[kevingranade]

Summary

SUMMARY: None

Purpose of change

The time spent in Travis builds has been climbing over time, delaying feedback for PR authors.
Also, we haven't reexamined our coverage of various elements in a while, so we had some missing elements.
While I was at it, more solidly enumerated just what we're testing in each build.

Describe the solution

Attempt to get in all the configurations we really care about in PR tests.
Max and Min versions of supported compilers.
A mix of Curses and Tiles/Sound builds.
A mix of Make and CMake builds.
Localization disabled.
The most popular (Tiles/Sound) Mac and Windows builds.
Address sanitization enabled on latest GCC and Clang.
Clang-tidy build.

Finally, round out all the remaining supported compiler versions only on master builds, along with the somewhat obscure Mingw-w64 Terminal build.

Replaced the no-install GCC 5.4 target with a Clang target, the goal is to get faster feedback to PR authors about obvious errors (rather than obscure ones that rely on specific compilers to trigger), and Clang builds are faster, even after accounting for time to install Clang. This also booted one build off the PR list, which means PR builds as a whole get significantly faster.

Alternatives

There are a basically unlimited number of ways to try and get full coverage here.

[kevingranade]

@jbytheway I'm a little confused how the clang 8 build ever worked.

[jbytheway]

@jbytheway I'm a little confused how the clang 8 build ever worked.

The addons: *clang8 refers back to a previously defined thing with addons: &clang8. Did you rearrange them so they were in the opposite order perhaps? That would have broken it.

[jbytheway]

I like the option of replacing the gcc 5.4 with a clang build. Good choice.

[kevingranade]

Aha, I'm a little fuzzy on the reference thing since the travis documentation doesn't really go into it, I think because it's just a YAML feature. I did reorder them.

My only concern here is that I'm leaving win/curses uncovered, and we JUST had a win/curses specific failure.

A more minor thing is I'm thinking about splitting off one more stage so single PRs can't monopolize the whole worker pool (5 workers), probably make it 1/3/4

I'm still unhappy about the number of builds, it's way too many.

[ZhilkinSerg]

Jenkins rebuild.

[kevingranade]

Failed mingw build, but it's unrelated to this issue.

src/debug.cpp: In function 'std::__cxx11::string windows_version()':
src/debug.cpp:1019:32: error: unknown option after '#pragma GCC diagnostic' kind [-Werror=pragmas]
 #pragma GCC diagnostic ignored "-Wcast-function-type"
                                ^
cc1plus: all warnings being treated as errors
Makefile:768: recipe for target 'objwin/tiles/debug.o' failed
make: *** [objwin/tiles/debug.o] Error 1