Benefit of Travis CI cache

* Mostly for POSIX (`linux`, `osx`)
* https://travis-ci.org/kangwonlee/cpp_study/jobs/470803028 would show

| | Before (#32.2 first) | After (#32.1 second) | B - A |
|-----|-----:|-----:|-----:|
| total | about 6 min | about 4 min | about 2 min |
| cache setup | 0.12 + 3.36 + 2.56 | 0.15 + 6.00 + 26.64 | -26.75 |
| `wget miniconda` | 0.73 | 0.01 | 0.72 |
| `bash miniconda.sh` | 25.43 | 0.01 | 25.42 |
| `update conda` | 34.64 | 8.96 | 25.68 |
| recipe | 183.45 | 56.57 | 126.88 |
| cache update | 118.93 | 5.37 | 113.56 |

Expect 2 min x 6 = 12 min CPU time save per triggering

Start with the linked Travis CI job and review the cache setup and timing table for the POSIX linux and osx builds. Compare the cache setup, miniconda, conda update, recipe, and cache update timings; done means the proposed cache approach is implemented and its build-time benefit is verified against the reported baseline.

Vom Indexierungsmodell aus dem Issue-Text verfasst.