ggml-cuda : reduce redundant expf calls in GDN KDA kernel

[simonxluo]

Summary

Pre-compute expf(g_t[i]) values once and reuse across multiple loops in the KDA (Key-Dependent Activation) branch of the GDN kernel.

Changes

• Add pre-computation loop for expf(g_t[i]) values before main computation
• Replace redundant expf() calls with pre-computed array g_exp[]
• Reduces expf calls from 2*S_v to S_v per token (e.g., from 256 to 128 calls for S_v=128)

Performance

Tested on AMD AI Max 395 (gfx1151, ROCm 7.2) with Qwen3.5-35B-A3B-heretic-v2-Q8_0:

Phase	Tokens	Before (t/s)	After (t/s)	Change
Prefill	64000	666.604	674.82	+1.23%
Decode	256	31.724	31.645	-0.25%

The decode variation (-0.25%) is within measurement noise. The prefill improvement (+1.23%) is consistent and beneficial for long-context scenarios.

Testing

• Model: Qwen3.5-35B-A3B-heretic-v2-Q8_0
• GPU: AMD AI Max 395 (gfx1151)
• ROCm: 7.2
• Parameters: --no-mmap -fa 1 -c 8192 -ngl 999
• Verified numerical correctness: output unchanged from baseline

Notes

• This is a pure optimization with no algorithmic changes
• Zero risk: computation is mathematically equivalent
• Particularly beneficial for long-sequence prefilling

AI Assistance Disclosure

Claude Code was used to identify the redundant expf() calls and suggest the pre-computation optimization strategy. The implementation was manually reviewed and tested.

Fixes #