ggml-hexagon: fix swiglu failure at test-backend-ops

[chraac]

Fixes some failure on the Hexagon swiglu/silu op and improves correctness of the GGML Hexagon backend.

Summary of changes

• Add overflow-guarded HVX primitives (hvx_vec_exp_fp32_guard, hvx_vec_inverse_fp32_guard) and use them in the Hexagon backend.
• Improve NaN/Inf handling in exp, inverse, and related helper func.
• Fix mistakes in the Hexagon silu / swiglu implementations (including handling of src1 and swapped/split variants).

Before

[SILU] NMSE = 3.457965465 > 0.000000100   SILU(type=f32,ne_a=[128,2,2,2],v=0): �[1;31mFAIL�[0m
[SILU] NMSE = 0.496767445 > 0.000000100   SILU(type=f32,ne_a=[5,7,11,13],v=0): �[1;31mFAIL�[0m
[SWIGLU] NMSE = 3894.832699597 > 0.000000100   SWIGLU(type=f32,ne_a=[128,2,2,2],v=0,swapped=0): �[1;31mFAIL�[0m
[SWIGLU] NMSE = 2.032236263 > 0.000000100   SWIGLU(type=f32,ne_a=[5,7,11,13],v=0,swapped=0): �[1;31mFAIL�[0m
[SWIGLU] NMSE = 531.844516626 > 0.000000100   SWIGLU(type=f32,ne_a=[128,2,2,2],v=0,swapped=1): �[1;31mFAIL�[0m
[SWIGLU] NMSE = 1.988691331 > 0.000000100   SWIGLU(type=f32,ne_a=[5,7,11,13],v=0,swapped=1): �[1;31mFAIL�[0m
[SWIGLU] NMSE = 1040.190893229 > 0.000000100   SWIGLU(type=f32,ne_a=[128,2,2,2],v=0,split): �[1;31mFAIL�[0m
[SWIGLU] NMSE = 0.493175916 > 0.000000100   SWIGLU(type=f32,ne_a=[5,7,11,13],v=0,split): �[1;31mFAIL�[0m

After

[SILU] NaN at index 231 (HTP0=-nan CPU=88.449669)   SILU(type=f32,ne_a=[128,2,2,2],v=0): �[1;31mFAIL�[0m
  SILU(type=f32,ne_a=[5,7,11,13],v=0): �[1;32mOK�[0m
[SWIGLU] NaN at index 122 (HTP0=nan CPU=-11446.431641)   SWIGLU(type=f32,ne_a=[128,2,2,2],v=0,swapped=0): �[1;31mFAIL�[0m
  SWIGLU(type=f32,ne_a=[5,7,11,13],v=0,swapped=0): �[1;32mOK�[0m
[SWIGLU] NMSE = 3.835742624 > 0.000000100   SWIGLU(type=f32,ne_a=[128,2,2,2],v=0,swapped=1): �[1;31mFAIL�[0m
  SWIGLU(type=f32,ne_a=[5,7,11,13],v=0,swapped=1): �[1;32mOK�[0m
[SWIGLU] NaN at index 216 (HTP0=nan CPU=-8444.154297)   SWIGLU(type=f32,ne_a=[128,2,2,2],v=0,split): �[1;31mFAIL�[0m
  SWIGLU(type=f32,ne_a=[5,7,11,13],v=0,split): �[1;32mOK�[0m

[chraac]

thought we can move this init out of the for loop below.

[chraac]

its safe to use the gcc ext since in htp we're using clang.

[chraac]

looks we should use src0->ne[0] here instead of dst->ne[0]

[chraac]

using xor here to fix when x < 0

[max-krasnyansky]

Looks good!

Thanks for the updates.
This PR also fixes #16854

[chraac]

This PR also fixes #16854

Just a heads-up: I've only fixed the opt_path == false path here. hvx_fast_sigmoid_f32 is still tricky because of the manual float exponent math causing NaNs, so I'll tackle that in a separate PR.

[max-krasnyansky]

This PR also fixes #16854

Just a heads-up: I've only fixed the opt_path == false path here. hvx_fast_sigmoid_f32 is still tricky because of the manual float exponent math causing NaNs, so I'll tackle that in a separate PR.

Yep. I saw that other tests are still failing.
But the Qwen3-0.6B output is definitely "fixed", as in it's actually sensible now :)