[ternary-f32] Phase 2: TernaryF32GemvNative SPI + reference ViewKernel + TernaryF32KernelPack

[michalharakal]

Part of #1136 (Phase 2 of 5). Pure commonMain in skainet-backend-api — the SPI, the portable reference kernel, and the dispatch pack. No native dependency; mirrors the existing BitNetGemvNative / TernaryKernelPacks seam.

Tasks

• TernaryF32GemvNative.kt — array-shaped native seam:

  public interface TernaryF32GemvNative {
      public val name: String
      /** out[o] = Σ_k in[k] · code(o,k) — scale NOT applied; sequential
       *  4-codes-per-byte BitNet payload. inputDim % 4 == 0. */
      public fun gemvPacked(
          activation: FloatArray, activationOffset: Int,
          weight: ByteArray, weightByteOffset: Int,
          inputDim: Int, outputDim: Int,
          out: FloatArray, outOffset: Int,
      )
  }

• TernaryF32GemvKernel.kt — portable reference ViewKernel: FP32 activation × BITNET_B1_58 weight, decode ((byte >> 2*(i%4)) & 3) − 1, fp32 accumulate, post-multiply by TernaryCodec.bitNetScale. keyFor() = the exact key matmul(FP32 dense contiguous × FP32/BITNET_B1_58 BLOCKED_ROW_MAJOR).
• TernaryF32KernelPack.kt — install(native: TernaryF32GemvNative?, capabilities, warn):
  • registers the native-backed view kernel under both the capability-free and capability keys (mirroring TernaryKernelPacks.install) only when native != null — without the native LUT kernel, the existing requantize→bitnet_gemv path remains the better portable fallback, so the exact key stays unregistered and behavior is unchanged
  • native view kernel: unwrap heap storage, loop rows for prefill (advance offsets by r*k / r*n), apply bitNetScale to outputs, fall back to the f32 reference (not BitNetGemvKernel) for non-heap storage
• Rebase the Phase-1 FFM object onto the SPI: FfmTernaryF32Gemv : TernaryF32GemvNative + an install(warn) helper in skainet-backend-native-cpu/jvmMain

Tests

• TernaryF32KernelPackTest (FakeNative, modeled on TernaryKernelPacksTest): exact-key selection beats the requantize path; pack absent → int8 path still serves; multi-row loop; scale application; install(null) warns and registers nothing
• TernaryF32GemvKernelTest: reference vs TernaryCodec.decodeBitNet + naive FP32 matmul on random tensors (exact equality on the codes-dot; tolerance only for summation order)