fix(models/deepseek-v3): correct MoE reshape for 3D forward input

[inureyes]

Follow-up hardening for #525 (round 2): the Kimi-VL port loads its real checkpoint (round-1 config fix #608) but aborted during the forward pass with [broadcast_shapes] Shapes (1,91,12288) and (1,91,6,1) cannot be broadcast (running -p "Describe this image in one sentence." --image tests/fixtures/test_image.png, after expanding 1 media placeholder to 64 image tokens).

Root cause

The crash is not in the MoonViT vision tower, connector, or merge; those produce a correct [1, 91, 2048] merged-embedding stream. It is the shared DeepSeek-V3 MoE backbone that Kimi-VL reuses as its text model.

SwitchGLU::forward in src/models/deepseek_v3.rs ended with reshape(&output, &[n_tokens, top_k, -1]), where n_tokens and top_k were read from indices.shape[0] and indices.shape[1]. That is only valid for a 2D [n_tokens, hidden] token stream. The text stack runs on 3D [batch, seq, hidden] states (attention reads batch/seq from the leading axes), so indices is [batch, seq, top_k]; the reshape then read n_tokens = batch = 1 and top_k = seq = 91 and folded the real top_k (6 = num_experts_per_tok) and hidden (2048) axes into one, producing [1, 91, 12288]. moe_weighted_sum then broadcast that against scores[..., None] = [1, 91, 6, 1] and aborted. Note 12288 = 6 * 2048 and 91 = 64 image tokens + prompt. It surfaces via Kimi-VL because kimi-vl-a3b is the first small DeepSeek-V3-architecture MoE exercised end to end here (full DeepSeek-V3 is too large to run locally).

What changed

• src/models/deepseek_v3.rs SwitchGLU::forward: drop the rank-assuming final reshape and return the already-squeezed [..., top_k, hidden], matching the proven deepseek_v2 sibling and the shared moe_weighted_sum contract. This is an identity for 2D input (the reshape was already a no-op there) and fixes the 3D path, so both the text-only and VLM embeddings paths are correct. deepseek_v32 has its own already-correct SwitchGLU and is unaffected.
• Adds two deterministic MoE regression tests (no checkpoint, no GPU): switch_glu_forward_preserves_rank_for_2d_and_3d and moe_block_3d_forward_matches_reference_shape.

Test plan

• cargo test --lib deepseek_v3 (17 pass, including the two new tests)
• cargo test --lib kimi_vl (16 pass, no regression)
• cargo clippy --lib --tests -- -D warnings (clean)
• cargo fmt