This PR is blocked by feat(profiler): add CPU-to-NPU correlation for ACLprof

【合入来源】

如有社区issue，请关联issue链接 请勿携带内部流程信息（需求链接、问题单、内部issue等）

• 需求
• 问题单
• issue/工单
• 重构优化
• 资料更新

【修改方案】

请描述修改内容的具体实现，涉及哪些组件之间进行交互，可以用1、2、3、...进行罗列
如果是需求或者重构类的PR，需要补充详细设计文档（说明上下游组件关系、时序图、类图、DFX能力等内容）

This PR adds opt-in diagnostics and hardware validation support for ACLprof CPU↔NPU correlation in the Kineto profiler backend. The existing correlation behavior remains unchanged when diagnostics are disabled.

1. Add opt-in ACLprof correlation diagnostics

   • Introduce diagnostic recording controlled by TORCH_NPU_ACLPROF_CORRELATION_DEBUG.
   • Keep diagnostics disabled by default so that normal profiling behavior is unaffected.
   • Record structured diagnostic events for correlation capture, task-queue processing, HostToDevice conversion, kernel resolution, matching, clock conversion, and final Kineto linkage.
   • Preserve diagnostic artifacts in JSONL format for post-run investigation.

2. Add correlation and lifecycle counters

   • Track Kineto correlations captured at enqueue time.
   • Track enqueue/dequeue task-queue marks and HostToDevice records.
   • Track converted launch records and successful clock conversions.
   • Track matching outcomes including FOUND, NOT_FOUND, and AMBIGUOUS.
   • Track final CPU↔NPU links.
   • Emit per-session summaries to make correlation behavior observable without changing production matching semantics.

3. Keep diagnostic work outside correlation-critical locking where possible

   • Defer diagnostic formatting and file I/O until after the correlation mutex is released.
   • Preserve the existing matching and linking logic while minimizing diagnostic overhead in correlation-critical paths.

4. Improve correlation failure observability

   • Record clock reference and conversion results.
   • Record candidate matching information and final match results.
   • Record kernel-resolution and linkage decisions so missing or ambiguous CPU↔NPU associations can be investigated from retained artifacts.

5. Add hardware-independent unit-test coverage

   • Verify diagnostics are inactive when debug mode is disabled.
   • Verify diagnostic state is reset correctly across profiling sessions.
   • Verify successful Kineto correlation capture and matching paths.
   • Cover negative and edge cases including ambiguous matching, missing matches, clock conversion failures, fallback behavior, diagnostic I/O failures, and profiling lifecycle transitions.

6. Add a manual hardware validation tool

   • Add tools/kineto_aclprof_correlation_validation.py for validation on real Ascend hardware.
   • Run two consecutive ACLprof profiling sessions and validate the complete correlation chain from captured CPU correlation IDs through converted HostToDevice launches to final NPU kernel links.
   • Validate clock conversion, matching results, final links, and session isolation.
   • Retain diagnostic and profiling artifacts when validation fails to simplify investigation.

7. Add correlation diagnostics documentation

   • Document prerequisites, invocation, environment configuration, PASS/FAIL criteria, diagnostic counters, and generated artifacts.
   • Clarify that the hardware validation script is a manual maintainer validation tool and is not registered as a CI test.

【资料变更】

请确认是否涉及资料变更。如涉及，需要在PR中体现，并简要说明修改内容。如不涉及，需填写“不涉及”

Documentation is updated.

Added docs/aclprof_correlation_diagnostics.md describing:

• how to enable ACLprof correlation diagnostics;
• how to run the hardware validation tool;
• required runtime prerequisites;
• automatically configured environment variables;
• PASS and FAIL criteria;
• interpretation of diagnostic counters;
• retained diagnostic and profiler artifacts for failure analysis;
• the manual, non-CI status of the hardware validation tool.

【接口变更】

请确认是否涉及跨代码仓或者客户面可见的接口变更。如涉及，需要详细说明接口以及对应的变更内容，同时需要在资料中体现。如不涉及，需填写“不涉及”

No cross-repository or customer-visible API changes.

The added diagnostic environment variables, diagnostic artifacts, and validation script are intended for profiler diagnostics and maintainer validation. Existing profiler APIs and the default profiling behavior are unchanged.

【功能验证】

说明测试场景，测试方法。如果本次测试方式与常规单元测试不同，请详细说明您的测试步骤
新增/变更内容是否已新增/适配UT测试用例看护，并补充测试自验证截图

The changes were validated with both hardware-independent unit tests and a manual hardware validation run on Ascend hardware.

1. Profiler unit tests

   • ACLprof correlation diagnostics and lifecycle scenarios are covered by profiler GTests.
   • Tests cover diagnostics enabled/disabled behavior, session reset, successful correlation capture, matching results, ambiguity handling, clock conversion failures, fallback paths, diagnostic I/O failures, and lifecycle behavior.
   • The profiler test suite passed with 33/33 tests.

2. Static/script validation

   • git diff --check passed.
   • python3 -m py_compile tools/kineto_aclprof_correlation_validation.py passed.
   • python3 tools/kineto_aclprof_correlation_validation.py --help passed.

3. Real Ascend hardware validation

   • The validation tool was executed as:

mkdir -p ./kineto_aclprof_cor_output

python3 tools/kineto_aclprof_correlation_validation.py \
    --output-dir ./kineto_aclprof_cor_output \
    --device 0

• Two consecutive ACLprof profiling sessions were validated.

• Each session produced:

  • converted_launches = 24
  • successful_clock_conversions = 24
  • FOUND = 24
  • NOT_FOUND = 0
  • AMBIGUOUS = 0
  • final_links = 24

• Both sessions completed successfully and the tool reported:

ACLprof correlation validation PASS

This confirms that CPU↔NPU correlation is correctly captured, converted, matched, linked, and reset across consecutive profiling sessions on real Ascend hardware.

【CheckList】

PR提交人对以下CheckList自检项进行全量自检，自检通过或不涉及，均修改 [ ] 为 [x]

• 代码注释完备，正确记录错误日志
• 代码实现进行了返回值、空指针等校验
• PR标题正确使用类型标签，如：feat、fix、refactor、docs、test等
• PR持续集成流水线（CI）执行通过，代码检查无异常