fix Fuzz constant OOM for CUDA (#8481)

## Summary

The Fuzz check has been failing for CUDA `compress_gpu` job for a very
long time now (at least a month).

After some investigation, it appears this is because ASAN will mprotect
a very large memory region that conflicts with some address space that
the CUDA driver wants to own, resulting in cudaCtx initialization
failing with OOMs like we've been seeing.

I've run a sample Fuzz workflow with this change and confirmed we are
able to get past context init now.

Signed-off-by: Andrew Duffy <andrew@a10y.dev>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: a10y

.github/workflows/fuzz.yml

@@ -200,6 +200,10 @@ jobs:
       fuzz_target: compress_gpu
       runner: gpu
       extra_features: "cuda"
+      # ASan mprotects the shadow gap PROT_NONE by default, which collides with the large
+      # virtual-address range the CUDA driver reserves and makes device init fail with an
+      # OOM-style error. protect_shadow_gap=0 frees that VA range for CUDA.
+      extra_env: "ASAN_OPTIONS=protect_shadow_gap=0"
       jobs: 1
     secrets:
       R2_FUZZ_ACCESS_KEY_ID: ${{ secrets.R2_FUZZ_ACCESS_KEY_ID }}