fix(layernorm): allow normalized_size==1 instead of error (#25896)

[Jonahcb]

Description

• Relax LayerNorm input validation to accept norm_size == 1; only reject zero-size (and less than zero).
  • Change norm_size <= 1 → norm_size < 1.
  • Update error text to “larger than 0”.
• File: onnxruntime/core/providers/cpu/nn/layer_norm_helper.h
• Add tests for the degenerate case (normalized extent is 1):
  • LayerNorm_NormSizeEqualsOne_Succeeds: with bias → output equals bias.
  • LayerNorm_NormSizeEqualsOne_NoBias: without bias → output all zeros.
• File: onnxruntime/test/contrib_ops/layer_norm_op_test.cc
• No compute-kernel changes; existing CPU/CUDA implementations already produce normalized zeros for norm_size == 1, yielding bias (if provided) as final output.

---

Motivation and Context

• Fixes a regression where LayerNormalization errored for norm_size == 1:

  Size of X.shape[axis:] must be larger than 1, got 1

• ONNX spec does not forbid norm_size == 1, and PyTorch returns zeros in this case; ONNX Runtime should align for backward compatibility with older models.
• Semantics for norm_size == 1: mean == x, var == 0 → normalized == 0 → output equals bias if provided, else 0.
• No behavior change for norm_size > 1.

Resolves: #25896

---

Files Changed

• onnxruntime/core/providers/cpu/nn/layer_norm_helper.h
  • Allow norm_size == 1; only reject norm_size < 1.
  • Update error message to “larger than 0”.
• onnxruntime/test/contrib_ops/layer_norm_op_test.cc
  • Replace failure case for size 1 with success tests (with/without bias).

[Jonahcb]

@kunal-vaishnavi This failure doesn't seem to be related to my edits. Are you able to assign reviewers to get this merged? Thank you!

[Jonahcb]

@snnn @fs-eire @skottmckay @yuslepukhin @kunal-vaishnavi It's not clear to me who is responsible for which part of the repository, but could you please assign reviewers to this so that this bug can be patched? Thank you!