llama/ggml: explicit signs, latent embeddings, grouped out_proj for the Hadamard contract

[bri-prism]

Follow-up to #119, stacked on hadamard-gguf-contract. That PR shipped the identity-sign contract; this one adds what a real trained checkpoint needs.

What

• Explicit sign vectors. sign_mode "explicit" carries per-width sign vectors in GGUF metadata (prism.hadamard.sign_widths plus a flattened sign_values). The loader materializes one F32 sign tensor per (width, buffer type) alongside the rotation, and the graph applies x' = H (s * x): an elementwise sign flip, then the existing blockwise FWHT hint matmul. Identity mode is unchanged and unknown modes still refuse to load.
• Inverse transform for latent embedding tables. Tensors listed in prism.hadamard.inverse_weight_names store latent (rotated) rows; the graph restores the primal basis right after the token-embedding lookup, h = s * (H z), so an embedding table can stay in the quantized latent format instead of a dequantized copy. The direct-embeddings input path is untouched.
• Grouped-order linear-attention out_proj. The converter's tiled V-head reorder permutes out_proj's input axis, which is the rotation axis of a folded latent; a post-fold column permutation breaks the blockwise correspondence and cannot be refolded without destroying the quantized codes. For folded out_proj tensors the converter now keeps the grouped V order and sets prism.hadamard.gdn_v_grouped; the runtime permutes the activation tiled-to-grouped (reshape, permute, cont) before the sign flip and rotation.
• Converter accepts schema v2 manifests (signs table, per-tensor roles), plus two missing template instantiations that only surface with the GNU linker.

Why

Trained Hadamard checkpoints use seeded sign flips and rotate every projection including the embedding and output head. Without these three pieces the contract could only represent identity-sign folds on plain matmuls.

Validation

• test-backend-ops -o MUL_MAT_HADAMARD: 14/14 on CUDA (SM90) and Metal against the CPU reference.
• End to end on a large hybrid-attention checkpoint packed with every projection folded (explicit signs, block 1024): logit KL divergence against the same model in unfolded F16 form lands at the measurement noise floor (mean KLD 3.7e-4, median 3.1e-4, same top-1 98.7%, mean ln PPL ratio -0.00005), i.e. the folded quantized pack is equivalent to the reference to the limit of the measurement.
• The activation-side transform stack measures at roughly 10% of GPU time on an H200 trace (FWHT 4.8%, sign multiplies 4.3%, permute copies about 2%), with obvious fusion headroom later.

[bri-prism]

Closing in favour of #121, which is the same work ported onto prism-v7.

This targets prism-v6, and the prism-v7 README now marks that branch as a stale mid-migration snapshot that should not be built from, so this cannot land as-is.

Nothing is lost. #121 carries all 18 commits from this branch and this one, rebased onto prism-v7, with the Metal FWHT kernels re-landed into the new ggml-metal/kernels layout since v7 moved the shaders out of the monolithic metal file. Verified there: 18/18 MUL_MAT_HADAMARD on CUDA (RTX 4090, 2/2 backends) and 17/17 on Metal (3/3 backends), and a folded checkpoint loads with 402 folded weights plus 1 inverse-lookup table and decodes at the same rate as on this branch.