Add model-comparison tooling and structured MMLU reporting

[Mikyx-1]

Implements step 1 of the series proposed in #1002.

Adds a serial baseline-versus-target runner with Markdown and JSON reports covering MMLU accuracy, correctness flips, full-vocabulary KL divergence, timing, memory usage, and optional cross entropy.

• Scores the first answer token using A–D label logits.
• Stores reference logits in JSON Lines with exact dataset, tokenizer, and prompt validation.
• Adds CMake/Bazel integration, regression tests, and usage documentation.
• Preserves configured inference settings.

Validation

• Clean CMake build.
• 12 C++ and 18 Python tests passed.
• C++ tests passed under AddressSanitizer and UndefinedBehaviorSanitizer.
• Completed an 83-question Gemma 3 270M self-comparison and cross-entropy evaluation.
• Verified rejection of mismatched reference prompts.

Existing autotuning produced one correctness flip during self-comparison. A temporary diagnostic using fixed MatMul candidates produced zero KL across five questions. Autotuner controls remain scheduled for step 2.

Bazel rules were updated but not executed because Bazel was unavailable.