#2554 tests: the golden setup configs name the image encoder as <VEXE>, so they compare on Linux too

Description

Same family as #1935 (setup tests that pass or fail depending on the PC), but not in its list.

**What it changes.** `tools/test_setup_golden.py`'s `normalize()` takes the engine's platform name out of a config
(`<EXE>`, #973) but not the image encoder's: `setup.VEXE` is `strata-vision.exe` on Windows and `strata-vision`
elsewhere. The five Unsloth UD-Q4_K_XL baselines with images were recorded on Windows, so on Linux the test failed
8 checks (`test_yes` x5, `test_enter_for_every_question` x3: `'<T>/engine/strata-vision'` != `'<T>/engine/strata-vision.exe'`).
`normalize()` now also turns the encoder's whole name into `<VEXE>`, with the same whole-name pattern as `<EXE>`, and
those five baselines say `<T>/engine/<VEXE>`.

**What it leaves alone.** `setup.py`, and every other baseline value (the diff of `test_setup_golden.json` is the
five `vision.exe` lines). The existing pattern's lookarounds are `[w.-]` (a literal `w`, probably meant as `\w`);
I kept them as they are for both names rather than change what `<EXE>` matches.

**How it was checked.** On Linux (Ubuntu, Python 3.14): `python3 tools/test_setup_golden.py` fails 8 on
`upstream/main`, OK with this change. `python3 tools/test_setup_golden.py --write` on Linux records a file identical
to the committed one. Windows was not run; with `setup.EXE`/`setup.VEXE` set to the Windows names, a reviewer checked
that `...\engine\strata-vision.exe` normalizes to `<T>/engine/<VEXE>` and that `strata.exe` does not match inside it.

🤖 Generated with [Claude Code](https://claude.com/claude-code)