ggml-wegpu: handle the buffer aliasing for rms fuse

[Constannnnnt]

Overview

This PR addressed an edge case of #21983. I load and run a model in the browser, and I met this error:

ggml_webgpu: Device error! Reason: 2, Message: Writable storage buffer binding aliasing found between [BindGroup "RMS_NORM_MUL"] set at bind group index 0, binding index 0, and [BindGroup "RMS_NORM_MUL"] set at bind group index 0, binding index 2, with overlapping ranges (offset: 5242880, size: 4096) and (offset: 5242880, size: 4096) in [Buffer "tensor_buf3"].
While encoding [ComputePassEncoder (unlabeled)].DispatchWorkgroups(1, 1, 1).
While finishing [CommandEncoder (unlabeled)].

As the error showed, it was associated with the buffer overlapping. I used a coding agent to analyze the logs:

• the previous inplace flag was only checking if mul_src and dst overlapped
• it missed a scenario where the rn_src overlapped with dst, i.e. (rn_src==dst), leading to this tensor_buf3 issue.

I reused the convention from the 'binary' shader:

inplace means src0 == dst. => rn_src == dst
overlap means src1 == dst. => mul_src  == dst
src_overlap means src0 == src1.

Additional information

I didn't run any benchmark tests and only tested the model behaviour in the browser.

Requirements

• I have read and agree with the [contributing guidelines](https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md)
• AI usage disclosure: Yes, I used coding agents for debugging.

[reeselevine]

Thanks for the fix. I tested it out and it looks like even with skip_validation turned off, there is nothing in test-backend-ops that would have caught this right now. Potentially other tests would, we actually should turn off skip_validation, I originally added it for performance natively but I think the impact is minimal. FlashAttention needs to handle some aliasing first though, which would be a good target for #22199.

@yomaytk fyi, if you're working on other fusion chains it might be good to check to see what other kind of buffer aliasing can occur and add a test for it in test-backend-ops if possible.

[yomaytk]

Thanks for the fix, this looks good to me. How about adding a test case in test-backend-ops and confirming that it passes without skip_validation in this PR? If not, I'm happy to follow up with a separate PR.

@yomaytk fyi, if you're working on other fusion chains it might be good to check to see what other kind of buffer aliasing can occur and add a test for it in test-backend-ops if possible.

Got it, thanks.