ggml-webgpu: simplify flash_attn shaders and refactor several WGSL shaders

[yomaytk]

Overview

This PR refactors several WGSL files and simplifies the flash attention WGSL files. The changes to each WGSL shader are as follows:

• concat.wgsl, rms_norm_mul.wgsl, row_norm.wgsl, solve_tri.wgsl, ssm_scan.wgsl: Remove unused params.
• conv2d.wgsl, conv2d_dw.wgsl, im2col.wgsl: Define the macros as the actual type for weights, input, and output buffer and remove some unnecessary functions.
• soft_max.wgsl: Remove the unused param ne. Also, binding numbers for each tensor buffer are computed sequentially instead of defining buffers with nested macros, since the nested macro structures aren't easy to read.
• flash_attn_decls.tmpl: This file contains the common definitions of params and types shared among flash attention WGSL shaders. Binding numbers for the tensor buffers are computed in the same way as in soft_max.wgsl.
• flash_attn_staging.tmpl: flash_attn_quant_staging.tmpl was renamed to this file, and load_(k|v)_tile_block for the float type is extracted from the three flash attention shaders.
• flash_attn.wgsl, flash_attn_tile.wgsl, flash_attn_vec_split.wgsl: Common code is extracted into flash_attn_decls.tmpl and flash_attn_staging.tmpl.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, I used AI to find the redundant code for the all wgsl shaders, and to refactor the several parts. Those changes are reviewed and the other parts (mainly for FA) were written by me.