Optimize UNet inference: enable LLAMAFILE SGEMM + fuse SpatialTransformer 1x1 convolutions

[Copilot]

UNet inference optimized from ~2463ms to ~1332ms per step (~1.85x speedup) on CPU with FP16 weights.

Enable LLAMAFILE optimized SGEMM by default

GGML_LLAMAFILE_DEFAULT was OFF, leaving llamafile's tinyBLAS (cache-blocked AVX2+F16C matrix multiply) unused. Set it ON before add_subdirectory(ggml). This alone accounts for ~1.7x of the speedup since MUL_MAT dominates (~172 GFLOPS across 282 ops in the graph).

Fuse SpatialTransformer 1x1 Conv2d with permutes

For SD1's !use_linear path, proj_in/proj_out are 1x1 Conv2d ops. The original flow:

Conv2d_1x1 → [im2col + mul_mat + permute + cont] → permute + cont → reshape → [transformer] → permute + cont → reshape → Conv2d_1x1 → [im2col + mul_mat + permute + cont]

Now both paths (linear and conv-1x1) share a unified flow — permute once, reshape, apply weight as a linear op via ggml_mul_mat on the reshaped 2D weight, then permute back once:

permute + cont → reshape → mul_mat(w_2d, x) → [transformer] → mul_mat(w_2d, x) → permute + cont → reshape

This eliminates 2× im2col + 2× permute+cont per SpatialTransformer (16 in SD1.5 UNet = 32 fewer data copies).

Minor: add get_params() accessor to GGMLBlock

Needed to access Conv2d weight tensors directly for the reshaped mul_mat path.