`paddle.static.nn.while_loop` with to_static(full_graph=True) segfaults when body captures an eager tensor

bug描述 Describe the Bug

Describe the Bug

paddle.static.nn.while_loop can crash the Python process when it is used
inside paddle.jit.to_static(full_graph=True) and the loop body is defined
outside the converted function while capturing an eager Tensor.

Minimal reproducible example

import paddle

cached_features = paddle.ones([2, 4], dtype="float32")


def continue_loop(step):
    return step < 2


def update_step(step):
    # A common pattern is to read cached state in a loop body. In this
    # reproducer the cached tensor is captured from eager mode.
    _ = cached_features + 1.0
    return step + 1


@paddle.jit.to_static(full_graph=True)
def run_loop():
    step = paddle.zeros([], dtype="int32")
    return paddle.static.nn.while_loop(continue_loop, update_step, [step])


print(run_loop())

Reproduction result

The program crashes the Python process:

FatalError: `Segmentation fault` is detected by the operating system.
SignalInfo: SIGSEGV (@0x18)
Segmentation fault (core dumped)
exit code: 139

Expected behavior

Paddle should not crash the process. If loop body functions that capture eager
Tensors are unsupported in this static-graph construction path, Paddle should
raise a Python/C++ error explaining the unsupported pattern.

Environment

PaddlePaddle: 3.3.1
Python: 3.13.13
OS: Linux 6.8.0-100-generic x86_64, glibc 2.35
GPU: 2*Nvidia 4090

其他补充信息 Additional Supplementary Information

No response