webgpu: Optimize depthwise conv2d

[qjia7]

PERF

To see the logs from the Cloud Build CI, please join either our discussion or announcement mailing list.

---

This change is 

[qjia7]

@kainino0x @lina128 @jinjingforever Please take a look, thanks.

In our last meeting, I ever said that the perf of depthwiseConv2d was almost 2x slower than webgl on Intel GPU. With this change, the perf is close between webgpu and webgl.
hand_detector:

DepthwiseConv2dNative	13.88

becomes

DepthwiseConv2dNative	8.28

This optimization is mainly for filter size 3x3 and stride size is 1x1. It's widely used for all of we tested models.
For such kind of depthwise conv2d, due to the stride size is only 1x1, there will be many data are repeated access in each channel tile. If we only calculate one output data in one invocation, we need to access 3x3 data in x and 3x3 data in filter. So to get a 4x4 output data, we need (3x3+3x3)* (4x4) = 288 times memory access. However, if we calculate 4x4 data in one invocation, we only need to access (3x6)*4 data in x and (3x3)*4 data in filter. So it needs (3x6 + 3x3) * 4 = 108 times, which is only half of previous one.

[lina128]

Thank you for the explanation and great perf improvement!

Reviewable status: complete! 1 of 1 approvals obtained (waiting on @jinjingforever and @kainino0x)