Validate inputs in GroupNorm and InstanceNorm - #3653

Proposed changes

Two cases where the norm layers accept bad inputs and return meaningless results instead of raising:

1. InstanceNorm on an input with no spatial dimensions returns all zeros. For a (N, C) or (C,) input the reduction axes tuple is empty, so mx.mean returns the input itself and mx.var returns 0, and the normalized output is exactly zero everywhere:

import mlx.core as mx
import mlx.nn as nn

nn.InstanceNorm(dims=8)(mx.random.normal((4, 8)))
# all zeros, no error

2. GroupNorm does not check that num_groups divides dims. With say num_groups=4, dims=6 the reshape can still succeed depending on the spatial size, and the groups end up mixing channel and spatial positions, so the statistics are garbage. The pytorch_compatible=True path floor-divides the group size and silently drops the intended grouping too. PyTorch raises for this at construction.

Both now raise a ValueError, same idea as #3615 did for SinusoidalPositionalEncoding. Added tests to the existing group norm and instance norm tests.

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)