remove useless pipeline buffer between cpu to stage 0

[chengtbf]

移除掉一个过渡设计：

Pipeline 并行下， CPU stage 0 到 GPU stage 0 之间插入过一个 buffer，且这 buffer = stage num * 2。

当时的想法是， cpu copy 到 gpu 的数据，可能被后向消费，如果没有 buffer ，则会阻碍 stage 0 的流水。

但实际情况是，我们一定会给 GPU stage 0 自动插入一个 identity（module 层级已经做好了这个 insert），而后向只可能消费这个 identity ，而不是 copy 到 gpu 的 regst。 而这个 identity 连向 backward 的边，一定会被识别为需要插入 buffer 的边，导致我们反向消费前向 identity 一定可以经过 buffer。 因此 前向 cpu 0 到 gpu 0 不需要再额外插入 buffer。

同时这个还会解决一个 bug，即 ： img gpu decoder ，会消费一个 cpu 上的 kTensorBuffer 数据（non POD 类型的，二级指针）。此时如果插入一个 identity buffer，则 buffer copy 这个就会出错。

[github-actions]

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/8484/

[github-actions]

Speed stats:

GPU Name: NVIDIA GeForce GTX 1080 

❌ OneFlow resnet50 time: 129.6ms (= 12956.6ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 145.5ms (= 14545.4ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.12 (= 145.5ms / 129.6ms)

OneFlow resnet50 time: 75.9ms (= 7593.4ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 85.6ms (= 8559.8ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.13 (= 85.6ms / 75.9ms)

OneFlow resnet50 time: 52.1ms (= 10426.9ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 56.3ms (= 11265.8ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.08 (= 56.3ms / 52.1ms)

OneFlow resnet50 time: 38.9ms (= 7781.9ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 43.5ms (= 8691.1ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.12 (= 43.5ms / 38.9ms)

OneFlow resnet50 time: 35.0ms (= 6993.3ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 40.5ms (= 8105.7ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.16 (= 40.5ms / 35.0ms)

OneFlow swin dataloader time: 0.272s (= 54.458s / 200, num_workers=1)
PyTorch swin dataloader time: 0.151s (= 30.121s / 200, num_workers=1)
Relative speed: 0.553 (= 0.151s / 0.272s)

OneFlow swin dataloader time: 0.076s (= 15.154s / 200, num_workers=4)
PyTorch swin dataloader time: 0.042s (= 8.436s / 200, num_workers=4)
Relative speed: 0.557 (= 0.042s / 0.076s)

OneFlow swin dataloader time: 0.040s (= 8.028s / 200, num_workers=8)
PyTorch swin dataloader time: 0.023s (= 4.611s / 200, num_workers=8)
Relative speed: 0.574 (= 0.023s / 0.040s)

❌ OneFlow resnet50 time: 144.9ms (= 14494.7ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 166.2ms (= 16622.6ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.15 (= 166.2ms / 144.9ms)

OneFlow resnet50 time: 95.8ms (= 9575.8ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 123.6ms (= 12364.2ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.29 (= 123.6ms / 95.8ms)

OneFlow resnet50 time: 69.6ms (= 13911.2ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 88.7ms (= 17734.8ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.27 (= 88.7ms / 69.6ms)

OneFlow resnet50 time: 56.2ms (= 11234.8ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 74.7ms (= 14932.0ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.33 (= 74.7ms / 56.2ms)

OneFlow resnet50 time: 51.2ms (= 10239.6ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 73.7ms (= 14743.1ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.44 (= 73.7ms / 51.2ms)

[leaves-zwx]

fix: https://github.com/Oneflow-Inc/OneTeam/issues/1459

[github-actions]

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/8484/