sycl: fix soft_max_f32 max reduction

[someoneinjd]

Overview

Fix the second-stage max reduction in the SYCL softmax kernel when a workgroup contains more subgroups than WARP_SIZE.

Previously, after each subgroup wrote its partial max to buf_iw[warp_id], the cross-subgroup reduction only loaded:

max_val = buf_iw[lane_id];

This only reduces the first WARP_SIZE partial maxima. For example, on Intel GPUs, WARP_SIZE can be 16 while the softmax workgroup size is 1024, producing nwarps = 64. In that case, partial maxima from buf_iw[16..63] were ignored.
If the true row max is in one of the ignored subgroups, softmax subtracts a too-small max value. For large attention logits this can overflow exp(x - max) to inf, eventually producing NaN.

Requirements

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

[arthw]

It's good job!

Thank you!

[jlionhan]

This PR also fixes a bug in VAE processing of stable-diffusion.cpp that caused fully white image to be generated.

Thank you for your contribution!

[someoneinjd]

This PR also fixes a bug in VAE processing of stable-diffusion.cpp that caused fully white image to be generated.

Thank you for your contribution!

Haha, I actually discovered this problem when running stable-diffusion.cpp on my Lunar Lake laptop as well. Glad this could help.

Once this PR gets merged into ggml, I will submit an issue to stable-diffusion.cpp requesting them to update ggml.