de, el: use $CommandOffset instead of string-length($Prefix) in navigate.yaml

[michaldziwisz]

Closes #740 point 1.

ec36e057 replaced string-length($Prefix)+1 with $CommandOffset, because
$NavCommand is always English (ZoomIn, MoveNext) while $Prefix is the
spoken, translated word. de and el were the last two files still carrying the
old formula; this ports the change to them and touches nothing else.

This is deliberately a no-op today, which is why it is a good moment to do it.
Both languages still use the untranslated zoom/move/read/describe as their
Prefix, so string-length("move") happens to equal the length of the Move
stem and all 18 branches match. The first translator who localizes those strings
loses every direction word, silently.

What changed

• 9 uses of substring($NavCommand, string-length($Prefix)+1) → $CommandOffset
  in each file;
• the 4 set_variables calls now also set CommandOffset (5 for Zoom/Move/Read,
  9 for Describe), same values as en.

Translations are untouched: every line in the diff is one of those two shapes.
After the change the offsets and the list of 9 branches (In, InAll, Out,
OutAll, Next, Previous, Current, LineStart, LineEnd) are identical to
en/navigate.yaml.

Verification

Checked on the engine rather than only on paper, because the interesting property
here is "nothing changes":

• Languages::de → 209 passed, 0 failed, and the same 209/209 with the
  change reverted. So the existing German tests do not discriminate this at all,
  which matches the report: there are no navigation tests for de or el.
• To get a real measurement I wrote four throwaway navigation tests for de on
  the existing init_nav / do_navigate_command harness (the shape used by
  en and nb), with the expected strings read out of the engine rather than
  written by hand. Result: 4/4 with the change, 4/4 without it — the no-op is
  confirmed on live output.
• Validity control, so those tests are not vacuous: translating the prefixes
  (zoom → zoomen, move → gehe) makes the OLD formula produce
  "zoomen; in zähler; 1" — the direction word is simply gone — while the new one
  produces "zoomen in; in zähler; 1". Same simulation over all 15 navigation
  commands: the old formula breaks 7/15 for German prefixes and 15/15 for Greek
  ones.

I kept those tests out of this PR to hold it to the two files I promised in the
issue. Happy to send them separately if you want de navigation pinned.

One thing I noticed while measuring

Greek has no tests running at all. tests/Languages/el/ exists with 11 files,
but el is not declared in tests/languages.rs, so cargo filters everything out
(0 passed; 5139 filtered out). Current spread: pl 625, hu 589, nb 583, en 580,
fr 518, ru 513, sv 503, fi 488, zh 290, de 209, pt 174, vi 19, id 16, el 0.
So for el this change rests on the structural check only. Say the word and I
will open a separate issue for that — it looked like an oversight rather than a
decision.

Point 2 of the issue (comparing $NavCommand directly, the way ru does) is not
part of this PR; it changes all languages, so it deserves its own decision.

[NSoiffer]

This looks good.

I need to look into the el tests. You are correct that it should be part of languages.rs. However, the el.rs file has all the lines commented out. Hopefully they all pass. I'll give them a try.