CUDA: support non-contiguous rows in L2_NORM

[MagicalFlames]

Overview

The L2_NORM CUDA kernel assumed contiguous rows: it read x[col] directly and was guarded by GGML_ASSERT(nb00 == ts0) plus ggml_is_contiguous_rows in supports_op. As a result, test-backend-ops cases with noncontig_rows=1 were reported as not supported.

This PR adds an inner-row stride to the kernel so it indexes elements via x[col*stride_col], passes the stride down from ggml_cuda_op_l2_norm, and relaxes the support check for L2_NORM.

Additional information

Testing with test-backend-ops test -o L2_NORM on an RTX 3050 (sm_86, CUDA 13.0) shows improvement from 20/20 tests (with 10 noncontig_rows cases skipped as not supported) to 30/30 passed, including all noncontig_rows=1 cases.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: No