Make the WebGPU backend pass test-llama-archs tests

[fairydreaming]

Overview

This PR fixes multiple issues that caused crashes and tests failures when running test-llama-archs with WebGPU backend.

Additional information

WebGPU fixes are:

• handling of src0/src1 overlap in GLU OP (happened in Phi-3 that uses split SwiGLU)
• handling of b/c overlap in SSM_SCAN OP (happened in Mamba)

Additionally WebGPU can't handle zero-length tensors (happened with MiniCPM3 and DeepSeek V3.2 due to default rope length being equal to key length - so nothing was left for nope part), so I added small tweaks in test-llama-archs to prevent them.

This is my first time touching WebGPU backend, so someone with more knowledge shall thoroughly review these changes.

Edit: after checking freshly added MiniMax M3 of course it also crashed in WebGPU due to weird issue - ggml_add_inplace() resulted in src0 and src1 being on WebGPU backend but dst being on CPU backend. I temporarily replaced it with ggml_add().

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - consulted possible fixes with GPT-5.6, then ignored its suggestions and did it my way