CUDA ArgMax on all -inf returns INT32_MAX sentinel value

Nobody has claimed this yet.

### Describe the issue

The CUDAExecutionProvider's ArgMax kernel returns `INT32_MAX` (2147483647) when ALL
elements along the reduction axis are `-inf`. The optimizer constant-folds this case
correctly (returns 0), but the runtime CUDA kernel produces an out-of-range sentinel.

### To reproduce

```python
from onnx import helper, TensorProto
import onnxruntime as ort
import numpy as np

SHAPE = [4, 12] # any [N, M] where M >= 2
v_t = helper.make_tensor("v_t", TensorProto.FLOAT, SHAPE,
[float('-inf')] * (SHAPE[0] * SHAPE[1]))

nodes = [
helper.make_node('Constant', inputs=[], outputs=['v_in'], value=v_t),
helper.make_node('ArgMax', inputs=['v_in'], outputs=['v_out'],
axis=-1, keepdims=0),
]

out_vi = helper.make_tensor_value_info('v_out', TensorProto.INT64, [SHAPE[0]])
graph = helper.make_graph(nodes, 'minimal', [], [out_vi])
model = helper.make_model(graph, opset_imports=[helper.make_opsetid('', 11)])

# CUDA optimized (default)
sess_opt = ort.InferenceSession(
model.SerializeToString(),
providers=['CUDAExecutionProvider', 'CPUExecutionProvider'],
)

# CUDA unoptimized (reference)
so_ref = ort.SessionOptions()
so_ref.graph_optimization_level = ort.GraphOptimizationLevel.ORT_DISABLE_ALL
sess_ref = ort.InferenceSession(
model.SerializeToString(),
providers=['CUDAExecutionProvider', 'CPUExecutionProvider'],
sess_options=so_ref,
)

opt = np.asarray(sess_opt.run(None, {})[0])
ref = np.asarray(sess_ref.run(None, {})[0])

print(f"opt (optimizer on): {opt}") # [0 0 0 0]
print(f"ref (optimizer off): {ref}") # [2147483647 2147483647 2147483647 2147483647]
assert not np.allclose(opt.astype(np.float64), ref.astype(np.float64),
atol=0.5, rtol=0.1), "BUG DID NOT REPRODUCE"
```

### Urgency

No. I found this by fuzzing testing.

### Platform

Linux

### OS Version

Ubuntu 5.4.0-162-generic

### ONNX Runtime Installation

Released Package

### ONNX Runtime Version or Commit ID

1.27.0

### ONNX Runtime API

Python

### Architecture

X64

### Execution Provider

CUDA

### Execution Provider Library Version

CUDA 13.0, Driver 580.76.05, GPU: NVIDIA RTX 3080 Ti

Start at the CUDAExecutionProvider ArgMax kernel entry point and compare its all--inf behavior with the optimizer's constant-folding result shown in the Python reproducer. Add or update coverage for a [4, 12] input filled with -inf, then run the CUDA ArgMax test path and verify the runtime returns the expected in-range result rather than INT32_MAX.

Written by the indexing model from the issue text.