[webgpu] Fix wrongly used u32 of WGSL

[axinging]

With this change, when the real coord < 0, 0.0 will be returned. However, without this change, when the real coord < 0, 0.0 will also be returned. This is why existing case can not catch this error.

Below is explanation for without this change. For example:

// Before this PR, coordCol  is of type u32.
let coordCol = outCol * uniforms.stride[1] + uniforms.dilation[1] * WCol - uniforms.pad[1];

When the real result of right side < 0, coordCol (of type u32) will be interpreted as a very big data, this big data will fail the check done by coordsInBounds4D, so 0.0 will be returned.

---

This change is 

[axinging]

@qjia7 @haoyunfeix @xhcao @gyagp , PTAL

[qjia7]

@kainino0x @lina128 take a look, thanks.

[lina128]

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

---

tfjs-backend-webgpu/src/kernels/conv2d_mm_vec4_webgpu.ts, line 312 at r1 (raw file):

 if (coordRow < 0) {

If coordRow < 0, is this an error? Why handle it silently?

[axinging]

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

tfjs-backend-webgpu/src/kernels/conv2d_mm_vec4_webgpu.ts, line 312 at r1 (raw file):

 if (coordRow < 0) {

If coordRow < 0, is this an error? Why handle it silently?

coordRow < 0 happens when padding is applied during computing convolution. This is intended behavior, can not report error.

[lina128]

Reviewable status: complete! 3 of 1 approvals obtained (waiting on @axinging)

---

tfjs-backend-webgpu/src/kernels/conv2d_mm_vec4_webgpu.ts, line 312 at r1 (raw file):

Previously, axinging (Xu Xing) wrote…

coordRow < 0 happens when padding is applied during computing convolution. This is intended behavior, can not report error.

Ah, right. If it's uint, then it will get a large number, which is incorrect. Got it. Thanks.