encoding: fix trim() to be O(n) instead of O(n^2)
`encoding.trim()` iterated over the possible lengths smaller than the
input and created a slice for each. It then calculated the column
width of the result, which is of course O(n), so the overall algorithm
was O(n). This patch rewrites it to iterate over the unicode
characters, keeping track of the length so far. Also, the old
algorithm started from the end of the string, which made it much worse
when the input is large and the limit is small (such as the typical 72
we pass to it).
You can time it by running something like this:
```
time python3 -c 'from mercurial.utils import stringutil; print(stringutil.ellipsis(b"0123456789" * 1000, 5))'
```
That drops from 4.05 s to 83 ms with this patch (and most of that is
of course startup time).
Differential Revision: https://phab.mercurial-scm.org/D12089
#require symlink
https://bz.mercurial-scm.org/1438
$ hg init
$ ln -s foo link
$ hg add link
$ hg ci -mbad link
$ hg rm link
$ hg ci -mok
$ hg diff -g -r 0:1 > bad.patch
$ hg up 0
1 files updated, 0 files merged, 0 files removed, 0 files unresolved
$ hg import --no-commit bad.patch
applying bad.patch
$ hg status
R link
? bad.patch