feat(compile): n-ary composite seam and the skainet.sdpa composite (#1343, part 2)

[michalharakal]

Stacked on #1348 (part 1). The diff includes its commits until #1348 merges; after that this branch is rebased onto develop and the draft flag dropped. Only the last commit (feat(compile): n-ary composite seam and the skainet.sdpa composite) is this PR.

Summary

Part 2 of #1343, stacked on part 1: the composite seam becomes n-ary and scaledDotProductAttention gets its stablehlo.composite.

• n-ary seam (sk.ainet.compile.hlo.lowering): HloDecomposition.emit(e, operands, operandTypes, resultType, params), LoweringSpec.lower(context, emitter, operands, operandTypes, resultType, params), ConversionContext.emitComposite with one %argN: T per operand in the private decomposition function and every operand type in the composite line; the symbol covers all operand types. HloDecomposition.unary { e, x, type, params -> … } and a unary lower overload keep GeluLowering/ErfLowering as short as before (byte-identical output).

• SdpaLowering.spec = LoweringSpec("scaleddotproductattention", "skainet.sdpa", attributes = scale · causal · window · rightContext · maskKind, version 1). When a target's OpGranularityPolicy keeps the op fused the converter emits

  %v5 = stablehlo.composite "skainet.sdpa" %arg0, %arg1, %arg2, %arg3
      {composite_attributes = {scale = 0.25 : f32, causal = true, window = 4 : i32, rightContext = 0 : i32, maskKind = "combined"},
       decomposition = @skainet_scaleddotproductattention_…, version = 1 : i32}
      : (tensor<2x4x8x16xf32>, tensor<2x2x8x16xf32>, tensor<2x2x8x16xf32>, tensor<2xi32>) -> tensor<2x4x8x16xf32>

  with the optional dense mask as a fifth operand; the decomposition function is built by the same SdpaLowering.emit as the inline form, from the operand types alone. scale is the resolved value (1/√headDim when the node left it at 0) so a vendor kernel can match it directly. Attribute keys are the node parameter names (camelCase), as skainet.gelu does, rather than the snake_case sketch in the issue.

Test plan

• CompositeEmissionTest: sdpa composite with four and five operands, attributes, private func.func with four arguments after the entry function, MlirParser entry = @main, MlirValidator clean, composite body ≡ inline decomposition modulo SSA names, gelu-only policy leaves SDPA inline; gelu/erf cases unchanged
• SdpaMaskRulesHloTest, SdpaGqaHloExportTest, DynamicShapeHloExportTest and the rest of compile-hlo green
• apiDump (skainet-compile-hlo.api: HloDecomposition.emit signature, Companion.unary, n-ary lower/emitComposite, SdpaLowering.OP/spec)
• scripts/pr-gate.sh --quick, full gate
• stablehlo-opt round trip (--stablehlo-legalize-composite-to-call) not run: the tool is not on the dev/CI machines; the body-equals-inline test is the stand-in.