Vulkan divide with truncate rounding mode doesn't match eager/portable

🐛 Describe the bug

Vulkan divide ops with rounding_mode="trunc" do not match eager/portable. When rounding mode is "floor" the output matches. It's not immediately clear if this is within expected numerical bounds, but enough values are off by 1 that it looks like the computation does not match. Feel free to close this issue if this is the expected behavior.

import torch
from executorch.backends.vulkan.partitioner.vulkan_partitioner import VulkanPartitioner
from executorch.exir import to_edge_transform_and_lower, EdgeCompileConfig, to_edge
from executorch.extension.pybindings.portable_lib import _load_for_executorch_from_buffer
from typing import Callable, List, Optional, Tuple, Union

class Model(torch.nn.Module):
    def forward(self, x, y):
        return torch.div(x, y, rounding_mode="trunc")
        
model = Model()
inputs = (
    torch.randn(1, 12),
    torch.randn(1, 12),
)
eager_outputs = model(*inputs)

ep = torch.export.export(model.eval(), inputs)
print(ep)
lowered = to_edge_transform_and_lower(
    ep,
    partitioner=[VulkanPartitioner()],
    compile_config=EdgeCompileConfig(_check_ir_validity=False)
).to_executorch()
print(lowered.exported_program())

et_model = _load_for_executorch_from_buffer(lowered.buffer)
et_outputs = et_model([*inputs])[0]

print(f"Inputs: {inputs}")
print(f"Eager: {eager_outputs}")
print(f"ET:    {et_outputs}")

Outputs:

Inputs: (tensor([[ 1.3996,  0.2905,  0.9381,  0.7877,  0.5824, -0.5655,  0.3260, -0.0476,
          0.2626, -0.3083, -2.1130,  0.5721]]), tensor([[-0.7415,  0.2069, -0.5097,  0.2503,  1.0457,  0.6018, -0.6957, -0.0889,
          1.3877, -0.1224, -1.4198,  1.6500]]))
Eager: tensor([[-1.,  1., -1.,  3.,  0., -0., -0.,  0.,  0.,  2.,  1.,  0.]])
ET:    tensor([[-2.,  1., -2.,  3.,  0., -1., -1.,  0.,  0.,  2.,  1.,  0.]])

Versions

Run on Meta internal master, Jul 3, fbcode/SwiftShader

cc @SS-JIA @manuelcandales @cbilgin

Start by running the provided reproduction with the VulkanPartitioner and compare truncating division against eager and portable outputs. Then trace the Vulkan lowering path for torch.div with rounding_mode="trunc"; done means Vulkan results match the eager and portable results while floor behavior remains correct.