test_gelu_operation fails on Apple Silicon (1 ULP f32 precision mismatch)

**Describe the bug**

`cargo test test_gelu_operation` fails on Apple Silicon (tested on M4 Mac Mini) with a 1 ULP difference in `gelu_erf`:

```
assertion left == right failed
left: [[0.0, 0.8413447], [1.9544997, 2.9959502]]
right: [[0.0, 0.8413448], [1.9544997, 2.9959502]]
```

The expected value `0.8413448` is the x86 rounding of `GELU(1.0) = x · 0.5 · (1 + erf(x / √2))` = `0.841344745...`. Apple Silicon rounds the same f32 value to `0.8413447`. Both are valid representations of the true value.

**Reproduction**

```
cargo test -p candle-onnx -- test_gelu_operation
```

**Expected behavior**

Tests should pass on all architectures.

**Suggested fix**

Use `to_vec2_round(z, 4)` instead of exact `to_vec2::` comparison, matching the pattern already used by `test_sin_operation` and `test_cos_operation`.

Run `cargo test -p candle-onnx -- test_gelu_operation` and inspect the `test_gelu_operation` test alongside the rounding pattern in `test_sin_operation` and `test_cos_operation`. Done means the GELU test accepts the valid 1 ULP f32 results and passes on Apple Silicon and x86.

Written by the indexing model from the issue text.