ExecuTorch correctness failure for upsample_bilinear2d_aa

[manuelcandales]

🕵️‍♂️ Detected with FACTO

🐛 Describe the bug

FACTO Generated meta_tuple:

['ArgType.Tensor torch.float32 (1, 1, 8, 1)', 'ArgType.Shape [3, 8]', 'ArgType.Bool False', 'ArgType.FloatOpt None', 'ArgType.FloatOpt None']

Incorrect output for the following input:

input = torch.tensor([[[[-98.5], [49.875], [17.125], [-46.5], [10.625], [-95.875], [-3.875], [-4.625]]]])
output_size = [3, 8]
align_corners = False
scales_h = None
scales_w = None
torch.ops.aten._upsample_bilinear2d_aa(input, output_size, align_corners, scales_h, scales_w)

mismatching output (2nd row doesn't match ATen's):

aten_output = torch.tensor([[[
          [ -8.4408,  -8.4408,  -8.4408,  -8.4408,  -8.4408,  -8.4408,  -8.4408, -8.4408],
          [-23.1339, -23.1339, -23.1339, -23.1339, -23.1339, -23.1339, -23.1339, -23.1339],
          [-24.7368, -24.7368, -24.7368, -24.7368, -24.7368, -24.7368, -24.7368, -24.7368]]]])
executorch_output = torch.tensor([[[
          [ -8.4408,  -8.4408,  -8.4408,  -8.4408,  -8.4408,  -8.4408,  -8.4408, -8.4408],
          [ -8.7243,  -8.7243,  -8.7243,  -8.7243,  -8.7243,  -8.7243,  -8.7243, -8.7243],
          [-24.7368, -24.7368, -24.7368, -24.7368, -24.7368, -24.7368, -24.7368, -24.7368]]]])

PR containing this correctness tests: #13968

Versions

nightly

cc @larryliu0820