Fix single-client reader parallel

[leaves-zwx]

根据 @ouyangyu 反馈，#6222 会引起 single-client 下，Benchmark 中的 resnet50 在 fp16 下精度下降。

根据 @CPFLAME 反馈，lazy 多卡数据读取有问题，会读取重复的数据。

推测是这里 lazy 的某些用法导致 data reader op 会走到 ddp 的分支里面去。

[chengtbf]

OFRecordImageClassificationDataReader 还未迁移到 Multi-Client 下，所以如果这里判断了 is Multi-Client，60 行：if (ctx->op_type_name() == "OFRecordReader") 的特判就可以删掉了？

[leaves-zwx]

主要是 OFRecordImageClassificationDataReader 调用 61 行的 get attr nd_sbp 就会报错了，因为不存在这个 attr。

[CPFLAME]

验证结果：lazy多卡数据读取已被修复

修复之前的实验现象见：bert对齐进展issue
用两个数据集分别跑了bert lazy和graph的四卡训练：

• 1.包含四个不同data part 数据，diffpart
• 2.包含同样的四个不同的data part数据, 但是顺序不一致，shuffle

数据的分布如下图：

以下为实验结果：loss基本对齐

[lixinqi]

我们有很多地方都在内部特判，其实很危险。最好是比较高层的位置做if，这样能降低圈复杂度

[github-actions]

Speed stats:

GPU Name: GeForce GTX 1080 

OneFlow resnet50 time: 127.6ms (= 6377.9ms / 50, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 140.3ms (= 7016.8ms / 50, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.10 (= 140.3ms / 127.6ms)

OneFlow resnet50 time: 74.1ms (= 3706.5ms / 50, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 84.0ms (= 4201.4ms / 50, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.13 (= 84.0ms / 74.1ms)

OneFlow resnet50 time: 47.2ms (= 2362.4ms / 50, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 58.4ms (= 2921.7ms / 50, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.24 (= 58.4ms / 47.2ms)

OneFlow resnet50 time: 39.5ms (= 1974.5ms / 50, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 48.2ms (= 2412.1ms / 50, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.22 (= 48.2ms / 39.5ms)

OneFlow resnet50 time: 44.2ms (= 2209.7ms / 50, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 38.3ms (= 1917.1ms / 50, input_shape=[1, 3, 224, 224])
❌ Relative speed: 0.87 (= 38.3ms / 44.2ms)

OneFlow resnet50 time: 151.0ms (= 7551.3ms / 50, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 155.3ms (= 7763.2ms / 50, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.03 (= 155.3ms / 151.0ms)

OneFlow resnet50 time: 97.5ms (= 4874.9ms / 50, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 99.3ms (= 4963.1ms / 50, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.02 (= 99.3ms / 97.5ms)

OneFlow resnet50 time: 78.0ms (= 3899.0ms / 50, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 81.8ms (= 4090.7ms / 50, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.05 (= 81.8ms / 78.0ms)

OneFlow resnet50 time: 75.4ms (= 3772.3ms / 50, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 73.0ms (= 3650.9ms / 50, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 0.97 (= 73.0ms / 75.4ms)

OneFlow resnet50 time: 70.8ms (= 3539.4ms / 50, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 58.8ms (= 2942.4ms / 50, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 0.83 (= 58.8ms / 70.8ms)