perf(vit): fuse ViT axial RoPE via ggml GGML_ROPE_TYPE_VISION kernel

[gabilan]

The Hiera ViT encoder applied 2D axial RoPE with a hand-rolled complex multiply (sam3_apply_rope): reshape + strided (real,imag) views + 4 muls

• sub + add + concat + cont, ~8 primitive ggml ops per Q/K per block, 64 calls per encode. On Metal this dominated the encode: ~828ms / ~38% of a ~2170ms 1008x1008 encode.

ggml already ships a single fused vision-RoPE kernel (ggml_rope_multi with GGML_ROPE_TYPE_VISION), but it computes rotary frequencies internally from theta+positions and pairs head_dim elements split-half (i, i+n_dims), whereas the model stores a precomputed freqs_cis table paired interleaved ((2i, 2i+1)). This bridges the two conventions:

• Reorder head_dim once from interleaved to split-half before the kernel (new[c]=old[2c], new[c+n_dims]=old[2c+1]) so the kernel rotates the same (real,imag) pair the stored table does.
• n_dims = head_dim/2, sections {n_dims/2, n_dims/2, 0, 0}: X-axis frequencies land on head_dim pairs 0..15 (position block 0), Y-axis on 16..31 (block 1), freq = base^(-c/16) -- identical to the stored compute_axial_cis layout.
• freq_scale interpolates global-block positions back to the WS x WS pretrain grid (WS/n_img_embd = 24/72 = pretrain_image_size/image_size); windowed blocks are native (1.0). This reproduces the stored table's scale_pos.

The head_dim reorder is deliberately NOT undone: RoPE touches only Q and K, and the attention score Q.K^T is invariant under a head_dim permutation applied consistently to both, while V (and the output basis) is untouched -- so the attention result is unchanged while saving the reorder-back copy.

Validation (real q8_0 model, Metal, cat.jpg):

• The live-computed rotary frequencies were checked element-wise against every block's stored freqs_cis table: max |stored - computed| = 1.19e-07 (f32 epsilon) across all 32 blocks, both windowed and global.
• Golden mask regression vs the pre-change build: point/box/text-detection masks agree at IoU 0.999986 / 0.999996 / 0.999992 (3-12 disagreeing pixels out of 1.44M, all on mask boundaries); confidence deltas ~6e-6.
• Full downstream test suite (395 assertions, 17 cases) passes unchanged.

Isolated encode wall-clock: 2173.5ms -> 1378.8ms on Metal (1.58x).