Fix reverse shape inference in LayerNorm

[sxjscience]

Fix #17654. Now, the user will see an error message if the in_channels does not match with the corresponding dimension in the input

After the PR, the following will raise an error

import mxnet as mx
from mxnet.gluon import nn
net = nn.LayerNorm(in_channels=10)
net.initialize()
net.hybridize()
out = net(mx.nd.ones((2, 11)))  # Trigger the error

Error:

MXNetError: MXNetError: Error in operator layernorm0_layernorm0: Shape inconsistent, Provided = [10], inferred shape=[11]

[sxjscience]

@zheyuye