I forgot that Go 1.11 and earlier don't know how to install Go modules
globally. Revert to the old go get line, which does the right thing for
most users not yet on modules.

For example, it should match "test.xn--8y0a063a" just like it matches
"test.联通".

Instead of doubling the size of the regexp by adding the punycode
version of every known TLD, simply match any valid punycode string which
follows "xn--". It's highly unlikely that this would cause false
positives.

Fixes #27.

It was a convenient way to obtain the input in chunks, so that the tool
could print urls incrementally without having to read the entirety of
the input at once.

Unfortunately, we failed to notie that bufio.Scanner has a hard limit on
the size of each "token". In our particular case, it meant that any
sequence of many thousands of input bytes without any whitespace could
make the tool error out.

Instead, use bufio.Reader, which grows a buffer to fit the data being
read. Go back to reading one line at a time, as it can only stop at one
specific byte like '\n', and not many of them like all whitespace
characters.

Fixes #28.

This was an unintended bug added ages ago. In c64477d, we modified
midChar to add \p{Po} to the character set. Unfortunately, we added |,
essentially ending up with [somechars|\p{Po}]. In other words, | was now
part of the set, while we only meant it as a regex "or" operator.

Fix this, and add a test.

Fixes #26.

Without it, it only matches the end of the entire input.

Fixes #28, again.

Some users were repeating the regexp compile, so make that more obvious.
Thanks to Kevin Burke for the idea.

While at it, update the go.mod language version.

Fixes #31.

This way, we match URLs within markdown emphasis like *this* or _this_
without the trailing character.

It's highly unlikely that URLs would genuinely want to end in either of
those characters. Asterisks are reserved, so they shouldn't be in the
path to begin with. And underscores are most common in the middle of the
path, and are already common to style plaintext at the start or end of
URLs.

Still allow them in the middle of URLs, because there's no reason not
to.

Fixes #29.