feat(rocke): add scalar optimization barrier APIs

[tenpercent]

ISSUE ID : AICK-2253

Kernel authors sometimes need an intermediate floating-point result to round before the next operation. Combining a multiply with an add or FP16 conversion can change that result.

Add IRBuilder.optimization_barrier(value) and rocke_b_optimization_barrier(builder, value). They return the same scalar value and type while preventing the compiler from combining operations across it. The implementation uses the existing inline-assembly operation; booleans and byte-sized values widen to i32 and then return to their original type.

This PR targets develop independently of the FP4 and NVFP4 work. It includes dedicated Python/C parity emitters, scalar bit-preservation and rounding tests, and documentation linked from the API index. The guide explains supported types, use of the returned value, and the difference from sched_barrier and memory/thread synchronization.

Validation:

• Fresh gfx1250 build: 24 GPU tests passed through COMGR and HIP, with no skips. Includes mixed lane predicates, arbitrary scalar bit patterns, multiply-add rounding, and FP32-to-FP16 rounding.
• Strict Python/C++ CPU tests: 36 passed, plus 25 subtests; 8 CTests passed.
• Python/C++ LLVM parity: 71 families / 774 configurations for each of LLVM 20, 22, and 23. One initial LLVM 22 RMSNorm emitter timed out; its focused retry passed without code changes.
• Repository hooks, relative-path guard, and diff check passed.

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🔎 pre-commit	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—
🤖 therock-pr-bot	✅ Pass	—

🎉 All checks passed! This PR is ready for review.

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

[therock-pr-bot]

🎉 All checks passed! This PR is ready for review.