[CI/Build] Repair CUDA test dependency updates

[yangzhuxinyzx]

Purpose

Audit and safely land the useful dependency updates from #304, #305, and #306 on the latest main without importing the incompatible chardet update from #307.

• Upgrade sentence-transformers 5.2.0 -> 6.0.0.
• Upgrade Pillow 10.4.0 -> 12.3.0.
• Upgrade Black 24.10.0 -> 26.5.1 and lock its new required pytokens dependency.
• Repair the shared CrossEncoder test runner to use the supported model_kwargs parameter (valid in both SentenceTransformers 5.2 and 6.0).
• Keep chardet at 5.2.0 because mbstrdecoder==1.1.3 requires chardet>=3.0.4,<6.

Audit notes

The original Dependabot PRs changed only one lock line each. Re-resolving the CUDA test lock showed:

• Bump black from 24.10.0 to 26.5.1 #306 omitted Black 26's required pytokens~=0.4.0 package.
• Bump chardet from 6.0.0.post1 to 7.6.0 #307 is unsatisfiable with the current mbstrdecoder dependency and was automatically resolved back to chardet 5.2.0.
• The existing test runner passed removed automodel_args to CrossEncoder; both SentenceTransformers 5.2 and 6.0 expose model_kwargs instead.

Test Plan / Result

• uv pip compile requirements/test/cuda.in -c requirements/cuda.txt -o requirements/test/cuda.txt --index-strategy unsafe-best-match --torch-backend cu130 --python-platform x86_64-manylinux_2_28 --python-version 3.12 — passed, 316 packages resolved.
• Import/signature smoke with SentenceTransformers 6.0.0 + Transformers 5.5.3 — passed.
• CrossEncoder signature smoke with SentenceTransformers 5.2.0 — passed.
• Pillow 12.3 image composition/drawing smoke — passed.
• Black 26.5.1 formatting smoke with pytokens==0.4.1 — passed.
• pre-commit run --files requirements/test/cuda.txt tests/conftest.py — all applicable hooks passed.

No GPU or end-to-end model test was run.

Supersedes the effective changes of #304, #305, and #306 after merge. #307 should remain open until its parent dependency permits chardet 6+.