ggml-webgpu: add support for NVFP4

[yomaytk]

Overview

This PR adds NVFP4 support to ggml-webgpu. The following table shows the performance of CISCai/gemma-4-31B-it-NVFP4-turbo-GGUF on M2 and NVIDIA Tesla-V100.

M2:

Model	Test	master (t/s)	PR (t/s)
gemma4 31B NVFP4	pp512	13.61 ± 0.78	44.06 ± 8.07
gemma4 31B NVFP4	tg128	0.90 ± 0.06	2.51 ± 0.01

Tesla-V100:

Model	Test	master (t/s)	PR (t/s)
gemma4 31B NVFP4	pp512	1.39 ± 0.00	30.38 ± 0.09
gemma4 31B NVFP4	tg128	1.04 ± 0.00	3.35 ± 0.04

Since WebGPU doesn't seem to provide direct access to the GPU instructions of the host device, NVFP4 likely offers no performance advantage over k-quants on WebGPU backend.
The primary motivation is ecosystem coverage — NVFP4 GGUFs are increasingly available on huggingface, and this ensures ggml-webgpu can run them rather than falling back to CPU backend.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, I used AI to generate some boilerplate code.

[reeselevine]

nice, I agree that if right now people are using nvfp4 on webgpu they should probably be pointed towards using a k-quant instead, but still nice to have some acceleration if for some reason they insist on using them :)