inplace update done via aliased outputs should have more strict checks

[jjsjann123]

# torch version: 2.7.0a0+git1335882
# cuda version: 12.8
# nvfuser version: 0.2.25+git6b0b17e
import torch
from nvfuser import FusionDefinition, DataType

def nvfuser_fusion_id0(fd : FusionDefinition) -> None :
    T0 = fd.define_tensor(shape=[4, 2], contiguity=[False, True], dtype=DataType.Float, is_cpu=False, stride_order=[0, 1])
    S1 = fd.define_scalar(0.00000, dtype=DataType.Double)
    T2 = fd.ops.gt(T0, S1)
    S3 = fd.define_scalar(0.00000, dtype=DataType.Double)
    T4 = fd.ops.where(T2, T0, S3)
    T5 = fd.ops.cast(T4, dtype=DataType.Float)
    T6 = fd.ops.set(T5)
    T7 = fd.ops.permute(T6, [1, 0]) # this alias shouldn't be allowed
    fd.add_output(T7, T0)
    fd.add_output(T7)

with FusionDefinition() as fd:
    nvfuser_fusion_id0(fd)

inputs = [
    torch.randn(10, dtype=torch.float32, device='cuda:0').as_strided((4, 2), (1, 5)),
]
o = fd.execute(inputs)

Thanks to @csarofeen for the chat.
Our code base handles aliases in a restricted manner. We assume that the alias source and target should have identical logical domain (shapes). The above code example should error out, since the output size doesn't match the input sizes. (2, 4) vs (4, 2).

[jjsjann123]

linking comment: #4028 (review)