ssm_conv: channels-major input mode to drop the delta-net transpose

[roberteg16]

Summary

Qwen3.5/3.6 hybrid models (Gated DeltaNet + MoE) fed ggml_ssm_conv a time-major
conv input [time, channels, seqs], while the delta-net projection produces
channels-major [channels, time, seqs]. The required transpose showed up in
prefill as a concat_non_cont kernel scaling super-linearly with the ubatch, and
after forcing it contiguous it remained as a ~221 ms CONT (transpose) kernel at
ub=4096.

This PR removes it in two steps:

1. delta-net: force contiguous conv-state concat input — materialize the
   transposed input so the conv-state concat takes the coalesced concat_cont path
   instead of concat_non_cont (which scaled super-linearly with the ubatch).
2. ssm_conv: add channels-major input mode; drop delta-net transpose — add an
   opt-in channels-major layout to ggml_ssm_conv
   (ggml_ssm_conv_channels_major, selected via op_params[0]=1): input is
   [d_inner, d_conv-1+n_t, n_s] (channels contiguous), output layout unchanged.
   build_conv_state keeps qkv_mixed channels-major, prepends the recurrent conv
   state along the time axis (concat dim 1), and calls the new entry point —
   eliminating the transpose entirely.

• Implemented on CPU and CUDA (short- and long-token kernels; coalesced smem load).
• Other backends' supports_op reject op_params[0]!=0 so the scheduler falls back to
  CPU; existing (time-major) callers are bit-identical.
• Shared by the qwen35, qwen35moe, and qwen3next graphs.

Measured impact (Qwen3.6-35B-A3B, prefill pp4096, gfx1151)

Max ubatch (-ub 4096 -b 4096 → M=4096) — per-commit breakdown

op	baseline	after commit 1 (contiguous concat)	after commit 2 (channels-major)
CONCAT	1166 ms	34 ms	34 ms
CONT (transpose)	10 ms	221 ms	10 ms
SSM_CONV	47 ms	47 ms	43 ms
total prefill	3825 ms	2925 ms	2717 ms

• Commit 1 removes the quadratic concat_non_cont blowup (1166 → 34 ms) but relocates
  the transpose cost into a 221 ms CONT kernel: −899 ms (−23.5%).
• Commit 2 eliminates that transpose (CONT 221 → 10 ms) with no regression elsewhere:
  a further −208 ms.
• Combined: 3825 → 2717 ms, −1108 ms (−29%).

Default batch (-ub 4096 -b 2048 → M=2048; the realistic config, since n_batch defaults to 2048 and n_ubatch is clamped to it)

op	baseline	this PR (channels-major)
CONCAT	283 ms	35 ms
CONT (transpose)	11 ms	11 ms
SSM_CONV	51 ms	43 ms
total prefill	2786 ms	2590 ms

At the smaller chunk the concat_non_cont cost is lower (the blowup is super-linear),
so the win is smaller but still meaningful: −196 ms (−7.0%), driven by CONCAT
283 → 35 ms with no added transpose.

Test plan

• test-backend-ops -o SSM_CONV — new channels-major cases pass (CUDA vs CPU reference), short and long-token paths.
• Perplexity healthy across the three archs sharing the helper: qwen3next (Qwen3.6-35B-A3B) 3.40, qwen35moe (Qwen3.5-35B-A3B) 2.18, qwen35 (Qwen3.5-0.8B) 16.28.
• Non-CUDA backends only exercised via the CPU fallback guard; not run on Metal/Vulkan/SYCL/etc.
• ggml-backend-meta.cpp (tensor-parallel split) still assumes time-major — untouched, out of scope.