webgpu : avoid crash when offset is not multiple of 4 in WebGPU ggml_backend_tensor_get() implementation

[fairydreaming]

Overview

WebGPU uses CopyBufferToBuffer() that requires size and offsets to be multiplies of 4 in its ggml_backend_tensor_get() implementation. While the implementation rounds size internally, offset was not rounded, so it crashed for offset values not equal to a multiple of 4.

Additional information

This fix adds a local offset to make sure the global offset passed to CopyBufferToBuffer() is a multiple of 4.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO

[fairydreaming]

@ggerganov WebGPU CI failed anyway because of #27960, so #28071 is needed too.

[CISC]

Bah, this again:
https://github.com/ggml-org/llama.cpp/actions/runs/33365844609/job/99406209639?pr=28045

[yomaytk]

LGTM, thanks!

[fairydreaming]

Bah, this again: https://github.com/ggml-org/llama.cpp/actions/runs/33365844609/job/99406209639?pr=28045

@CISC Should be OK now, but needs re-approval

[yomaytk]

Bah, this again:
https://github.com/ggml-org/llama.cpp/actions/runs/33365844609/job/99406209639?pr=28045

@fairydreaming FYI, We format the WebGPU C++ code with clangd version 22 or later. But the EditorConfig error looks like it's already been resolved :)

[fairydreaming]

Bah, this again:
https://github.com/ggml-org/llama.cpp/actions/runs/33365844609/job/99406209639?pr=28045

@fairydreaming FYI, We format the WebGPU C++ code with clangd version 22 or later. But the EditorConfig error looks like it's already been resolved :)

@yomaytk Yeah, the reason was DOS/Windows EOL characters that GitHub added when I accepted @CISC review changes