No backward pass for `LayerNorm` if tensor is contiguous and remove_mean is true

This is similar to #2168, but the root cause appears to be different:

* In #2212, a new `candle_nn::ops::layer_norm` function was added, which uses `apply_op3_no_bwd`.

* In #2213, `LayerNorm`'s forward pass was modified to call `candle_nn::ops::layer_norm` when the tensor is contiguous and `remove_mean` is true.

Trace LayerNorm's forward path for contiguous tensors with remove_mean enabled, then inspect candle_nn::ops::layer_norm and its apply_op3_no_bwd usage. Reproduce the missing backward pass and compare with the behavior discussed in issues #2168, #2212, and #2213. Done means this LayerNorm path supports backward computation and has a regression test.

Written by the indexing model from the issue text.