fix(cuda_core): force UTF-8 mode for stubgen-pyx pre-commit hook

Windows CI (and local Windows developers) were regenerating .pyi files
with corrupted non-ASCII characters (e.g. Kepler(TM) -> Kepler mojibake).
stubgen-pyx's `include`-directive expansion reads .pxi files via
`Path.read_text()` with no explicit encoding, which falls back to the
platform's default (cp1252 on Windows) instead of UTF-8. `_device.pyx`
pulls in several .pxi files with UTF-8 docstring content ("Kepler(TM)",
"(in {mu}s)"), so this only surfaces there.

Invoke stubgen-pyx via `python -X utf8 -m stubgen_pyx` so Python's UTF-8
mode is forced regardless of platform locale, matching what the old
toolshed/run_stubgen_pyx.py wrapper did via PYTHONUTF8=1 before it was
removed. Putting this in .pre-commit-config.yaml (rather than only in
CI) means local Windows developers get the same fix.

This also revert the temporary --show-diff-on-failure diagnostic flag
added to track down the mismatch.

Root cause filed upstream: jon-edward/stubgen-pyx#41 (encoding fix was
incomplete - it covered top-level .pyx/.pxd reads but missed the
include-expansion path in file_parsing.py).

Author: mdboom

.github/workflows/ci.yml

@@ -531,7 +531,7 @@ jobs:
         shell: bash
         run: |
           set -euxo pipefail
-          SKIP=lychee pre-commit run --all-files --show-diff-on-failure
+          SKIP=lychee pre-commit run --all-files
 
   checks:
     name: Check job status

.pre-commit-config.yaml

@@ -83,7 +83,11 @@ repos:
 
       - id: stubgen-pyx-cuda-core
         name: Generate .pyi stubs for cuda_core
-        entry: stubgen-pyx cuda_core/cuda --continue-on-error --include-private
+        # -X utf8 forces Python's UTF-8 mode so file reads/writes don't fall
+        # back to the platform's default encoding (e.g. cp1252 on Windows),
+        # which otherwise corrupts non-ASCII characters (e.g. "™", "μ") in
+        # `include`d .pxi files. See jon-edward/stubgen-pyx#41.
+        entry: python -X utf8 -m stubgen_pyx cuda_core/cuda --continue-on-error --include-private
         language: python
         files: ^cuda_core/cuda/.*\.(pyx|pxd)$
         pass_filenames: false