add autocudagraph

[DrRyanHuang]

Description

Add autocudagraph decorator for automated CUDAGraph acceleration of forward and backward passes.

🌟 Key Features & Implementation

src/paddlefleet/cudagraph.py：核心基础设施

• 核心上下文 (CUDAGraphContext)：基于 dataclass 统一管理静态 Tensor Buffer、Fwd/Bwd 句柄与 PyLayer Runner。
• 泛型 IO 解析 (get_tensors / set_tensors)：支持递归遍历嵌套结构（Tensor / list / tuple / dict）提取与写回。
• 全自动捕获流 (autocudagraph)：支持 warmup -> capture -> replay 完整生命周期。
• 内存优化：前反向分别捕获为独立 CUDAGraph，共享同一 unique memory pool。
• 动态路由 & Fallback：引入 dispatch_key_fn，超出 max_graphs 上限时自动无缝回退 Eager 模式。
• 梯度累积保障：自动快照与恢复 nn.Layer 参数梯度，防止 dummy backward 污染真实的跨步累积梯度。
• 反向传播桥接：基于 paddle.autograd.PyLayer 正确路由输入张量梯度与权重参数梯度。

🧪 Test Coverage

tests/single_card_tests/test_autocudagraph.py

Test Class	覆盖点
TestPureFunctions	纯函数、多输入、stop_gradient 混合场景
TestOOPAndSubmodules	OOP 调用、权重指针漂移校验、多实例隔离
TestConfigurations	dispatch_key_fn 动态图分派、max_graphs Fallback 机制
TestComplexStructures	嵌套 dict/list IO、no_grad 上下文对齐
TestFatModel	多层 FFN 模拟、参数梯度端到端对齐
TestAdvancedMechanics	跨步梯度累积正确性、AMP 混合精度支持
TestEdgeMemoryOps	Tensor slice/stride 等复杂内存视角操作
TestGraphLimitationsAndFeatures	Dropout 随机性保持（兼容 Philox RNG）
TestDispatchNumber / TestNoGrad	max_graphs 缓存边界限制、no_grad 下的对齐

📝 Test Plan

• 单测全过：单卡 GPU 执行 test_autocudagraph.py，所有 Case 通过。
• 数值对齐：不同 warmup_steps 下，CG 与 Eager 模式前反向 100% 绝对一致 (rtol=0, atol=0)。
• Fallback 鲁棒性：输入组合超 max_graphs 限制时，正常执行不崩溃，平滑回退至 Eager 模式。
• 梯度安全：多步梯度累积下，累积梯度免受 Capture 阶段内部 Backward 的污染与清零。

[DrRyanHuang]

TODO: 记录一个 Paddle 的 BUG，很离谱，这里如果不带 .cpu()

            grads_eager.append(model_eager.head2.weight.grad.clone())

每次跑到固定位置，eager 的 grad 就回变成全 0！待排查具体原因

[codecov-commenter]

Codecov Report

❌ Patch coverage is 98.11321% with 3 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@94e7941). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
src/paddlefleet/cudagraph.py	98.11%	1 Missing and 2 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop     #896   +/-   ##
==========================================
  Coverage           ?   98.11%           
==========================================
  Files              ?        1           
  Lines              ?      159           
  Branches           ?       33           
==========================================
  Hits               ?      156           
  Misses             ?        1           
  Partials           ?        2           

Flag	Coverage Δ
coverage_combine	98.11% <98.11%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
src/paddlefleet/cudagraph.py	98.11% <98.11%> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.