CUDA: non-contiguous (RMS) norm support

[JohannesGaessler]

This PR adds CUDA support for non-contiguous input tensors for (RMS) norm.

[slaren]

Just making a note here, we cannot keep adding parameters like to every op test case, it makes them much more complex. This will need to be refactored at some point and replaced with a generic way to create non-contiguous views for the op parameters.

[ggerganov]

@JohannesGaessler Apply this patch to fix the Metal build:

diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
index 3ae4bbdd1..0a264be37 100644
--- a/ggml/src/ggml-metal/ggml-metal.m
+++ b/ggml/src/ggml-metal/ggml-metal.m
@@ -1206,11 +1206,11 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
         case GGML_OP_GROUP_NORM:
             return has_simdgroup_reduction;
         case GGML_OP_RMS_NORM:
-            return has_simdgroup_reduction && (op->ne[0] % 4 == 0 && ggml_is_contiguous_1(src0);
+            return has_simdgroup_reduction && (op->ne[0] % 4 == 0 && ggml_is_contiguous_1(op->src[0]));
         case GGML_OP_ARGMAX:
             return true;
         case GGML_OP_NORM:
-            return has_simdgroup_reduction && (op->ne[0] % 4 == 0 && ggml_is_contiguous_1(src0);
+            return has_simdgroup_reduction && ggml_is_contiguous(op->src[0]);
         case GGML_OP_ROPE:
             {
                 const int mode = ((const int32_t *) op->op_params)[2];