fix: handle multi-dim activations in LoKr forward

[undefeaterpeter]

Problem

Applying a LoKr LoRA (lokr_w1/lokr_w2 factors) to a model whose linear layers run on 3-D activations aborts the process with SIGABRT. Crash stack:

ggml_abort
ggml_reshape_3d
ggml_ext_lokr_forward
LoraModel::get_out_diff
MultiLoraAdapter::forward_with_lora
Linear::forward
Krea2::KreaAttention::forward

Root cause

The non-conv branch of ggml_ext_lokr_forward assumes a 2-D activation [q, batch] and takes batch = h->ne[1]. Krea2::KreaAttention (and other batched attention paths) feed linears 3-D activations [features, L, N]; whenever N > 1, ggml_nelements(h) = q·L·N ≠ vq·uq·L, so the first split reshape's element-count GGML_ASSERT fires and aborts. With N == 1 every reshape happens to pass, which is why LoKr works on many models and the gap went unnoticed.

Fix

• batch = ggml_nelements(h) / q_actual — fold all trailing dims into the batch (the existing reshapes already assume a contiguous activation).
• Reshape the result back to [up·vp, ne[1], ne[2], ne[3]] instead of flat 2-D, so forward_with_lora's ggml_add_inplace adds the diff element-wise onto the linear output rather than depending on a shape that no longer matches.
• The conv branch already batches on ne[3] and is unchanged. The batch dim stays outermost through every reshape/transpose in the pipeline, so folding L·N and restoring at the end is index-exact; the Kronecker row convention (out = u·vp + p) is unchanged.

Verification

• Repro: Krea-2 Q4_1 GGUF + a LoKr LoRA (full lokr_w1/lokr_w2 factors, OneTrainer/diffusers export) — 100% SIGABRT at the stack above before the fix.
• After: the same model + LoRA stack renders to completion (Metal backend, Apple M4); output is coherent, confirming the restored layout is index-exact and not merely non-crashing.
• Built and render-tested on the commit before the recent tokenizer changes; src/model/adapter/ is identical between that commit and current master.

🤖 Generated with Claude Code — reviewed and submitted by the repo owner per the AI-assisted contribution policy.

[undefeaterpeter]

Withdrawn.