This also includes a refactor of the hugofs package and its usage.
The motivation for that is:
The Afero filesystems are brilliant. Hugo's way of adding a dozen of global variables for the different filesystems was a mistake. In readFile (and also in some other places in Hugo today) we need a way to restrict the access inside the working dir. We could use ioutil.ReadFile and implement the path checking, checking the base path and the dots ("..") etc. But it is obviously better to use an Afero BasePathFs combined witha ReadOnlyFs. We could create a use-once-filesystem and handle the initialization ourselves, but since this is also useful to others and the initialization depends on some other global state (which would mean to create a new file system on every invocation), we might as well do it properly and encapsulate the predefined set of filesystems. This change also leads the way, if needed, to encapsulate the file systems in a struct, making it possible to have several file system sets in action at once (parallel multilanguage site building? With Moore's law and all...)
Fixes#1551
It is a fairly costly operation:
```
benchmark old ns/op new ns/op delta
BenchmarkScratchGet-4 109 31.6 -71.01%
benchmark old allocs new allocs delta
BenchmarkScratchGet-4 0 0 +0.00%
benchmark old bytes new bytes delta
BenchmarkScratchGet-4 0 0 +0.00%
´´´
This uses the Emoji map from https://github.com/kyokomi/emoji -- but with a custom replacement implementation.
The built-in are fine for most use cases, but in Hugo we do care about pure speed.
The benchmarks below are skewed in Hugo's direction as the source and result is a byte slice,
Kyokomi's implementation works best with strings.
Curious: The easy-to-use `strings.Replacer` is also plenty fast.
```
BenchmarkEmojiKyokomiFprint-4 20000 86038 ns/op 33960 B/op 117 allocs/op
BenchmarkEmojiKyokomiSprint-4 20000 83252 ns/op 38232 B/op 122 allocs/op
BenchmarkEmojiStringsReplacer-4 100000 21092 ns/op 17248 B/op 25 allocs/op
BenchmarkHugoEmoji-4 500000 5728 ns/op 624 B/op 13 allocs/op
```
Fixes#1891
We can of course skip reading the entire byte slice again and again.
This was a slip in the original implementation; functionally the same,
but is slightly faster, esp. for larger data sets with many shortcodes:
```
benchmark old ns/op new ns/op delta
BenchmarkReplaceShortcodeTokens-4 15505 14753 -4.85%
benchmark old allocs new allocs delta
BenchmarkReplaceShortcodeTokens-4 1 1 +0.00%
benchmark old bytes new bytes delta
BenchmarkReplaceShortcodeTokens-4 3072 3072 +0.00%
```
