feat: add x_attention_v2 operator use pypto as compile backend.

[Yangxu-China]

Description

Add the x_attention_v2 PyPTO Pro kernel to the xllm NPU kernel package, along with its JIT test suite.

This PR introduces a new DSL sub-package xllm/python/kernels_npu/pypto/ (parallel to the existing tilelang/ and triton/ sub-packages) containing the x_attention_v2 operator — a beam-decode flash attention kernel for Ascend A5 (Ascend950), implementing shared KV + unshared KV dual-path merged attention with online softmax combine.

The kernel is written in the PyPTO Pro DSL (pypto_pro.language) and launched via the @pl.jit bracket-launch path. A Python host wrapper (x_attention_v2) computes tiling (core split) and launches the @pl.jit kernel (x_attention_v2_kernel).

Key additions:

• xllm/python/kernels_npu/pypto/__init__.py — pypto DSL sub-package entry (license + docstring only, not registered in kernels_npu/__init__.py, consistent with tilelang/ and triton/)
• xllm/python/kernels_npu/pypto/x_attention_v2.py — @pl.jit kernel + Python host wrapper (1892 lines, self-contained)
• tests/python/test_x_attention_v2_jit.py — 144 parametrized test cases (3 GQA configs × 2 batch sizes × 3 beam sizes × 4 KV lengths × 2 unshared lengths) with self-contained golden cache mechanism
• .gitignore — entries for tests/python/golden_cache/ and tests/python/msprof*/

Verification:

• 144/144 tests pass on Ascend950 — 13.6s with golden cache present, ~186s on first run without cache (CPU reference computation)
• JIT compilation is cached per-process per-shape (first call ~2.4s compile, subsequent ~16ms cached)
• msprof performance: total 6981us across 144 cases, avg 48.5us per case

Related Issues

N/A — new operator integration, no tracking issue yet.

Change Type

• Bug fix
• New feature
• Performance improvement
• Refactor
• Documentation
• Test
• Build or CI

Pull Request Checklist

Thank you for contributing to xLLM. Before requesting review, please make sure the following items are complete.

PR Title and Commit Messages

• The PR title and each commit message follow the xLLM commit format: <type>: <subject>.

Allowed types: feat, bugfix, docs, test, refactor, chore, style, revert, perf, model, build, release.
The subject should use clear English, start with a verb, include at least 4 words, and end with ..

Pre-commit Checks

• I have installed pre-commit by running pip install pre-commit or an equivalent command.
• I have installed the hooks with pre-commit install.
• I have run pre-commit run --all-files and fixed any reported issues.

If you are unsure how to set up pre-commit, see the pre-commit documentation.

Self Review

• I have self-reviewed the code according to .agents/skills/code-review/references/custom-code-style.md, especially code written or assisted by AI.
• I have rebased this PR onto the latest main branch.

Build and Test Coverage

• Tests have been added or updated as needed.
• CUDA: python setup.py build test has passed on a CUDA machine.
• NPU: python setup.py build test has passed on an NPU machine.
• MLU: python setup.py build test has passed on an MLU machine.

Reviewer Notes

• pypto_pro is a CANN-bundled package (installed at /home/developer/Ascend/cann-9.2.0/python/site-packages/pypto_pro/), not declared as a pip dependency — same model as tilelang (conditionally installed via CANN setup). The kernel file imports it lazily inside the host wrapper, so the module is importable without pypto_pro installed.
• The golden_cache/ directory (144 .pt files) is gitignored — the test generates cache on first run if absent (~186s for all 144 CPU reference computations), subsequent runs load from disk (~14s).
• msprof*/ directories are also gitignored — profiling results are not committed.
• The kernel source (x_attention_v2.py) is self-contained with zero external path dependencies — all imports are standard library, pypto_pro, or torch.
• The .gitignore change adds two entries for test artifacts.
• Its operator prototype is consistent with the existing x_attention operator. It only supports the Ascend950 hardware and uses PyPTO as the backend compiler.
• Good performance gains are observed across all 144 test cases within the test scope. Lab test results show that its average performance reaches 1.4× of the original x_attention Ascend950 implementation.

[zhang-minchao]

这个不需要提交吧？