fix: preserve text in typed-content checkpoint templates

[Jaeyeong-CHOI]

Summary

Granite Vision 3.2 selects its standalone multimodal template, but that template silently drops plain-string message content. The loader now compares one string probe with one typed-text probe and stores a compatibility flag only when the typed form preserves the marker.

Typed ChatMessage renders for affected checkpoints use a borrowed typed-text view in apply_inner. Raw JSON requests, explicit template overrides, default processors, and string-compatible templates keep their existing content contracts.

What changed

• Add selection-time detection with conservative error handling and no per-request probing.
• Serialize affected messages directly from borrowed role/text fields without an intermediate JSON tree or copied strings.
• Add eight focused regressions with pinned Granite Vision 3.2 and Granite 4 fixtures, plus English and Korean technical reports.

Validation

• Combined CUDA graph targeted tests: 8 passed, 0 failed (text_wrap_tests); 9,105 unrelated mlxcel tests and 160 surgery tests were filtered out.
• Repository contract gate: version, kernel, llama compatibility, ROCm overlay, 123 Python tooling tests, and 49 WebUI fixtures passed.
• Real mlx-community/granite-vision-3.2-2b-4bit revision 351831ab5218b1f66bbfdfd57bd8f20d81ce8baf on NVIDIA GB10/CUDA: CLI answered Paris in 8 tokens; server string and typed-text requests both answered Paris with 55 prompt tokens and 8 completion tokens. The frozen baseline reproduced the 47-token prompt and unrelated greeting.
• Implementation review, security rereview, and static finalization found no remaining issues.
• Full workspace test and Clippy suites, hosted CI, ROCm inference, and macOS inference are outside this issue unit; the parent integration gate owns broad validation.

Reports

• TECHNICAL_REPORTS/2165-typed-chat-text-20261013.en.md
• TECHNICAL_REPORTS/2165-typed-chat-text-20261013.ko.md

Closes #2135

Fixes #2135