spec : add DFlash2 support and IQ1 pairing

[Jackson57279]

Overview

Adds native DFlash 2 on this IQ1 branch (port of ggml-org/llama.cpp ggml-org#27342 / ggml-org#27310, without NVFP4 / d2t / backend-sampler hunks that this tree does not have).

DFlash 2 still uses --spec-type draft-dflash. Detection is selector_top_k > 0.

• Two-tap dynamic conv around attn/FFN
• In-graph top-k selector lattice, CPU greedy walk
• Encoder fused into KV inject (one llama_decode, no encode + D2H)
• M-RoPE drafts: degenerate rope sections + 4-row inject positions

IQ1 pairing:

• Target: IQ1_XS / IQ1_XXS / IQ1_XXXS
• Draft: Q8_0 or Q4_K. If a DFlash draft is IQ1-quantized, selector / conv / fc.weight stay Q8_0
• AVX2 vec-dot for IQ1_XS and IQ1_XXS (same pattern as IQ1_XXXS)

Additional information

Does not rebase onto master. Master already has DFlash 2; this ports it onto codex/iq1-quants-training-yarn.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - implementation assistance; review the graph, selector walk, quant overrides, and AVX2 dots before merge