failures with supported OPs in test_decomp

[yuchengliu1]

🐛 Describe the bug

• AssertionError: Get None or [] without decomp
  "test_comprehensive_nansum_xpu_bool",
  "test_comprehensive_nansum_xpu_int",
  "test_comprehensive_nansum_xpu_uint8",
  "test_quick_nansum_xpu_bool",
  "test_quick_nansum_xpu_int",
  "test_quick_nansum_xpu_uint8",
  "test_comprehensive_nn_functional_binary_cross_entropy_with_logits_xpu",
• RuntimeError: "avg_pool2d_xpu" not implemented for 'Long'
  "test_comprehensive_nn_functional_avg_pool1d_xpu_int64",
  "test_comprehensive_nn_functional_local_response_norm_xpu_int64",
• RuntimeError: Difference from float64 is larger with decomposition nll_loss2d_backward.default than original on output 0. Original max diff: 1.1224856321843946e-05, Decomp max diff: 1.9292721803156054e-05 atol = 1e-07
  "test_comprehensive_nn_functional_nll_loss_xpu_float16",

Versions

c6981a2