[Web] Using ceil() in shape computation is not yet supported for MaxPool

### Describe the issue

I want to run a simple CNN using `onnxruntime-web` with WebGPU (or WebGL) to get some runtime acceleration and I cannot do this due to the following error:
```
ERROR Error: using ceil() in shape computation is not yet supported for MaxPool
at Array.Jh (ort.all.min.mjs:3702:627)
at jo.computeKernel (ort.all.min.mjs:4204:10470)
at Object.Xb (ort.all.min.mjs:4204:17298)
at 851205 (ort-wasm-simd-threaded.jsep.mjs:45:422)
at mc (ort-wasm-simd-threaded.jsep.mjs:92:259)
at ort-wasm-simd-threaded.jsep.wasm:0x1224637
at ort-wasm-simd-threaded.jsep.wasm:0x122535a
at ort-wasm-simd-threaded.jsep.wasm:0x10de910
at ort-wasm-simd-threaded.jsep.wasm:0x23b86c
at ort-wasm-simd-threaded.jsep.wasm:0x7155da
```

It's a pretty standard operation and I'm surprised that it's not supported. When do you plan to add the support for it?

### To reproduce

1. Create and export neural network to ONNX:

Example code

```python
import torch
from torch import nn

class Net(nn.Module):
def __init__(self):
super().__init__()
self.conv1 = nn.Conv2d(in_channels=3, out_channels=64, kernel_size=3)
self.maxpool = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True)

def forward(self, x):
x = self.conv1(x)
x = self.maxpool(x)
return x

def convert_to_onnx(model, example_input, opset_version=12):
# Prepare the model for export
model.eval()
# Set the name of the ONNX file
onnx_file_name = "net.onnx"
# Export the model to an ONNX file
torch.onnx.export(model,
example_input,
onnx_file_name,
export_params=True,
opset_version=opset_version,
do_constant_folding=True,
input_names = ['input'],
output_names = ['output'],
dynamic_axes={'input' : {0 : 'batch_size'}, # variable length axes
'output' : {0 : 'batch_size'}})

```

2. Load the network in JS/TS and use `webgpu` or `webgl` (both fail):
```typescript
import * as ort from 'onnxruntime-web/all';
ort.env.wasm.wasmPaths = 'https://cdn.jsdelivr.net/npm/onnxruntime-web@dev/dist/';

let model = await ort.InferenceSession
.create('./net.onnx'
,{executionProviders: ['webgpu'] }
);

function createRandomTensor() {
const size = [1, 3, 320, 320];
const values = new Float32Array(size.reduce((a, b) => a * b));

for (let i = 0; i < values.length; i++) {
values[i] = Math.random();
}

const tensor = new Tensor("float32", values, size);
console.log(tensor);
return tensor;
}

const feeds: Record = {};
feeds[model.inputNames[0]] = createRandomTensor();

const outputData = await model.run(feeds);
```

3. When the runtime is initialized **without webgpu**, like this:
```typescript
let model = await ort.InferenceSession
.create('./net.onnx')
);
// ... rest of the code
```
**it works fine.**

### Urgency

Not urgent, but utilizing WebGPU / WebGL would be beneficial for pretty standard CNNs.

### ONNX Runtime Installation

Other / Unknown

### ONNX Runtime Version or Commit ID

onnxruntime-web 1.19.0-dev.20240601-217b66fd85

### Execution Provider

'webgl' (WebGL), 'webgpu' (WebGPU)

Start with the WebGPU and WebGL execution paths for the MaxPool operator and the reported ceil() shape-computation error in ort.all.min.mjs. Reproduce the model with PyTorch and ONNX Runtime Web 1.19.0-dev, comparing webgpu or webgl against the working default runtime. Done means the same ceil_mode MaxPool model initializes and runs successfully on the affected web execution providers.

Written by the indexing model from the issue text.