[ONNX FE] Use MAX bias mode for L2 norm to guard zero-vector inputs

[gblong1]

Details

• Switched the L2 branch of lp_norm from BiasMode::ADD to BiasMode::MAX, so the epsilon bias now floors the denominator (sqrt(max(sum_of_squares, eps))) instead of being added to it (sqrt(sum_of_squares + eps)).
• The MAX mode only affects the zero-norm case; for any nonzero norm sum(x^2) >> eps, so the output is bit-identical to the unbiased computation and normal embeddings are not perturbed by the guard.
• Prevents division-by-zero / NaN for zero-vector inputs during L2 normalization without altering results for typical inputs.
• One-line change (1 insertion, 1 deletion); no signature or API changes — l2_norm already supported the BiasMode::MAX path.

Tickets:

• CVS-189394

AI Assistance:

• AI assistance used: yes
• Used to analyze CVS-189394 and to test the impact of this change
• Used to create the testcase

[Copilot]

Pull request overview

This PR updates the ONNX Frontend’s internal lp_norm helper so that L2 norm uses a flooring epsilon bias (sqrt(max(sum(x^2), eps))) rather than an additive bias (sqrt(sum(x^2) + eps)), aiming to stabilize normalization behavior around zero-vector inputs.

Changes:

• Switch lp_norm(..., p_norm==2) from BiasMode::ADD to BiasMode::MAX when computing the L2 denominator.

[gblong1]

@mvafin - any feedback on this one? it should be a simple review.