models: remove redundant ggml_cont calls now supported by CUDA unary kernels

[saady789]

Overview

This is a follow-up to #29781. That PR added support for unary operations on non-contiguous tensors in CUDA. With that support now upstream, this PR cleans up redundant ggml_cont / ggml_cont_2d allocations and their associated TODO comments across affected models.

Removing these intermediate copies simplifies the compute graphs and avoids unnecessary memory bandwidth overhead during the forward pass.

Changes

• src/models/delta-net-base.cpp:
  • Removed redundant ggml_cont on g_last before ggml_exp in build_delta_net_chunking.
• src/models/qwen3next.cpp:
  • Removed ggml_cont_2d on gate before ggml_sigmoid in build_layer_attn.
  • Removed ggml_cont on b before ggml_sigmoid in build_layer_attn_linear.
  • Removed ggml_cont_2d on gate before ggml_sigmoid in graph_mtp::graph_mtp.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO this was manual cleanup

[CISC]

Ah, not quite ready yet:
https://github.com/ggml-org/llama.cpp/actions/runs/37878431717/job/113696555416?pr=30198#step:8:113

[CISC]

Also, there are more that are not marked with TODO, f.ex. in dflash:

llama.cpp/src/models/dflash.cpp

Line 340 in 3b75c63

cat_conf = ggml_sigmoid(ctx0, ggml_cont(ctx0, ggml_view_2d(ctx0, base, 1, n_blocks, base_stride, 0)));

[saady789]

Also, there are more that are not marked with TODO, f.ex. in dflash:

llama.cpp/src/models/dflash.cpp

Line 340 in 3b75c63

cat_conf = ggml_sigmoid(ctx0, ggml_cont(ctx0, ggml_view_2d(ctx0, base, 1, n_blocks, base_stride, 0)));

Removed the redundant ggml_cont call in src/models/dflash.cpp.

Regarding src/models/qwen3next.cpp: kept ggml_cont_2d in place because the subsequent ggml_mul invokes ggml_mul_impl, which asserts GGML_ASSERT(ggml_can_repeat(b, a)). Without standard contiguous layout here, the repeat/broadcasting check fails. Verified the graph execution passes on Tesla T4.