Unet compare eager and graph test

[Ldpe2G]

新增 unet eager 和 graph 输出对比测试

为了排除 fused_multi_head_attention_inference 原子操作的影响，attention 中先注释掉 _fused_forward
相关讨论见：https://github.com/Oneflow-Inc/OneTeam/issues/1816#issuecomment-1342214821

测试运行方式：

python3 -m unittest discover -v -s tests/ -p 'test_models_unet_oneflow.py'  -k 
 UNet2DConditionModelTests.test_compare_eager_graph_output

测试思路

首先全部注释掉 tests/test_models_unet_oneflow.py 文件开头的所有环境变量设置，然后通过逐个开启，对比输出最大的 diff，来确定是哪个优化导致 eager 和 graph 输出不一致的。

os.environ["ONEFLOW_MLIR_CSE"] = "1"
os.environ["ONEFLOW_MLIR_ENABLE_INFERENCE_OPTIMIZATION"] = "1"
os.environ["ONEFLOW_MLIR_ENABLE_ROUND_TRIP"] = "1"
os.environ["ONEFLOW_MLIR_FUSE_FORWARD_OPS"] = "1"
os.environ["ONEFLOW_MLIR_GROUP_MATMUL"] = "1"
os.environ["ONEFLOW_MLIR_PREFER_NHWC"] = "1"
os.environ["ONEFLOW_KERNEL_ENABLE_FUSED_CONV_BIAS"] = "1"
os.environ["ONEFLOW_KERNEL_ENABLE_FUSED_LINEAR"] = "1"
os.environ["ONEFLOW_KERENL_CONV_ENABLE_CUTLASS_IMPL"] = "1"
os.environ["ONEFLOW_KERENL_FMHA_ENABLE_TRT_FLASH_ATTN_IMPL"] = "1"
os.environ["ONEFLOW_KERNEL_GLU_ENABLE_DUAL_GEMM_IMPL"] = "1"
os.environ["ONEFLOW_CONV_ALLOW_HALF_PRECISION_ACCUMULATION"] = "1"
os.environ["ONEFLOW_MATMUL_ALLOW_HALF_PRECISION_ACCUMULATION"] = "1"