Add fused QK-Norm + RoPE kernel for Qwen3

[zhangxin81]

Summary

Add a fused Triton QK-Norm + RoPE path for Qwen3 and wire it into the optional Qwen3 monkey patch path.

This PR:

• adds a new fused qk_norm_rope Triton op with forward/backward
• exposes a functional wrapper and a Qwen3 attention forward replacement
• adds an opt-in qk_norm_rope flag to apply_liger_kernel_to_qwen3
• adds correctness tests and a profiling script
• updates monkey-patch coverage and the README patching table

Testing Done

• Hardware Type: NVIDIA H100 80GB HBM3 (from the provided benchmark report)
• run make test to ensure correctness
• run make checkstyle to ensure code style (targeted local ruff/import/format checks on changed Python files)
• run make test-convergence to ensure convergence

Additional validation completed locally:

• imported liger_kernel successfully in a scratch Python 3.12 venv
• ruff check passed for all changed Python files
• ruff format --check passed for all changed Python files
• py_compile passed for the new Python files

Note: full make test / convergence runs were not executed in this shell environment because the default local Python environment did not have the CUDA/PyTorch stack installed.

Benchmark Summary

From the supplied performance report (report file intentionally not included in this PR):

Correctness

• 55 tests passed
• The fused path keeps the RMSNorm -> RoPE chain in fp32 and is measured against an fp32 gold reference
• For bf16, the fused path was reported as not worse than the unfused baseline and often more accurate because it avoids extra rounding

Performance (bsz=4, seq_len=2048, H100 80GB)

BF16

• qwen3_0.6b: forward 20.3x, full 7.2x, memory 577M -> 242M (58% reduction)
• qwen3_4b/8b: forward 20.2x, full 10.8x, memory 1089M -> 402M (63% reduction)
• qwen3_14b: forward 19.3x, full 13.2x, memory 1345M -> 483M (64% reduction)
• qwen3_32b: forward 20.4x, full 16.7x, memory 2113M -> 723M (66% reduction)

FP32

• qwen3_0.6b: forward 10.5x, full 7.2x, memory 770M -> 483M (37% reduction)
• qwen3_4b/8b: forward 10.6x, full 10.5x, memory 1287M -> 803M (38% reduction)
• qwen3_14b: forward 10.5x, full 9.9x, memory 1570M -> 964M (39% reduction)
• qwen3_32b: forward 10.5x, full 12.0x, memory 2434M -> 1444M (41% reduction)

Profiler summary

For the qwen3 4B/8B bf16 case, the report states the unfused path spent ~84.8 ms self CUDA time across many elementwise/cast kernels, while the fused path reduced that to ~10.2 ms dominated by the fused Triton forward/backward kernels.

Notes for reviewers

• The benchmark script is included in the PR.
• The benchmark/report markdown file is intentionally excluded, per request.
• The Qwen3 fused path is opt-in via qk_norm_rope=False by default.