ggml-webgpu: fix minimax-m3 crash on webgpu by adding F16 REPEAT

[yomaytk]

Overview

This PR fixes the error of webgpu ci error for minimax-m3 introduced by #24908.

minimax-m3 graph calls ggml_tensor * bmx = ggml_repeat_4d(ctx0, ...); followed by ggml_add_inplace(ctx0, bmx, ...), but existing webgpu backend doesn't have F16 repeat, so it crashes with Bus error, because the GPU ADD op ends up with an inplace destination on CPU buffer. So it fixes the error by adding F16 repeat support.

Additional information

The root cause seems to be in ggml scheduler, so I'll investigate it and follow up on it in separate PR.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, to diagnoise and write the code. code is reviewed by me.

[nikwen]

I didn't see your PR earlier, so I spent a few hours investigating this bug and came to the same solution. Implementing F16 REPEAT fixes the issue.

The crash happens because tensor->buffer->buft->iface.get_name(tensor->buffer->buft) is "CPU" here. We incorrectly cast tensor->buffer->context to ggml_backend_webgpu_buffer_context *. The program then crashes when we try to return ctx->buffer because we attempt an invalid memory access during the copy constructor of wgpu::Buffer.

[fairydreaming]

Tested this, it fixes the problem with crashing test-llama-archs minimax-m3 test.

[reeselevine]

There seems to be a similar potential bug in ggml scheduling that causes a few other models in test-llama-archs to fail too: #25931 (comment)

I noticed that the output buffer for a SET_ROWS operation is on the CPU, despite the input buffers being on the GPU.

[nikwen]

The minimax-m3 tests were disabled in #26415. They should probably be re-enabled after this PR was merged.

(I've hit the open PR limit so can't open another PR myself right now.)