Proposed changes

Fixes #3138
Bug: Large 3D convolutions ($M \times K > 2^{32}$) were causing silent data corruption due to 32-bit integer overflow in the naive_unfold_Nd kernel index calculation.
The Fix: Explicitly cast gid's to size_t before multiplying by filter_size to ensure 64-bit arithmetic for the output offset.
Verification: Verified locally (macOS 26.2, MLX 0.30.7, Apple M2) with a script very similar to the one described in the issue. Except using float16, and T=2500, H=64, W=64, C=16, to be just on the edge of 2^32, and my system's memory limit for a single buffer (some 9.5GB).
Result before change: M*K= 4,423,680,000 > 2^32 max_diff=152.884216 corr=0.956153 [FAIL]
Result after change: M*K= 4,423,680,000 > 2^32 max_diff=0.031334 corr=1.000000 [OK]

Didn't add a new test for this because it would need to allocate at least 2^32 * sizeof(type) to reproduce.

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)