[BUG] mx.var / mx.std wrong for offset data

☑️ I understand it is strictly prohibited to use AI to write issues.

Describe the bug
For data with a large mean relative to its spread, MLX variance is wrong by hundreds of percent while NumPy is exact:

x = 1e6 + 10 * noise      (10,000 float32 samples, true var = 100)
mx.var(x)   = 589.69      (+489%)
np.var(x)   = 97.53
mx.std(x)   = 24.28       (want ~10)
mx.mean(x)  off by ~22    (mean = 1e6, rel err 2e-5)

To Reproduce

Include code snippet

import mlx.core as mx, numpy as np
mx.set_default_device(mx.cpu)
rng = np.random.RandomState(0)
noise = rng.standard_normal(10_000).astype(np.float32)
x = (1e6 + noise * 10).astype(np.float32)
print(float(mx.var(mx.array(x))))     # 589.686  (true ~100)
print(float(np.var(x)))               # 97.527
print(abs(float(mx.mean(mx.array(x))) - 1e6))   # ~22  <- the poison

Start by running the Python reproduction with mx.var, mx.std, and mx.mean on the generated offset data, comparing each result with NumPy. Trace the MLX implementations behind these entry points and verify that the corrected results remain accurate for large-mean, small-spread float32 inputs.