feat(ops)!: exact (erf) GELU and erf op (#1342, part 1)

[michalharakal]

Summary

Part 1 of #1342: erf as a first-class op and an exact GELU, carried as op metadata so DIRECT, OPTIMIZED and the StableHLO export all see the same choice.

• TensorOps.erf(tensor) and gelu(tensor, approximate: Boolean = true). The default stays the tanh form, byte-for-byte, so no existing model changes numerically; approximate = false is 0.5·x·(1 + erf(x/√2)), the form BERT-class encoders and the Laya decision models were trained with. Measured on the kernel: the two differ by at most 4.73e-4 at |x| ≈ 2.70 (the issue's "≈ 2e-3" was an overestimate; the test pins the real band).
• CPU: ErfKernel (Abramowitz–Stegun 7.1.26 in double, |err| < 1.5e-7, the same expression order as the Laya host workaround so the scalar path reproduces it bit-for-bit) plus a JVM Vector API lane kernel (VectorOperators.EXP + Horner) with a single-precision scalar tail of identical operation order; scalar-vs-vector parity ≤ 1e-6 across SPECIES±1 tails. Drive-by: DefaultCpuOps.gelu printed a literal ${tensor.dtype} in its error message; the stale @InProgress(... task-ops.md) reference is gone.
• Metadata: GeluOperation(parameters["approximate"]) / ErfOperation in TensorOperations.kt; the KSP tracing wrapper records approximate next to the shape attributes (gelu/erf joined the unary list in MethodAnalyzer), so TraceToGraphBuilder puts it on the node; ComputeGraphExecutor dispatches on it; geluBackward picks the exact derivative Φ(x) + x·φ(x) from it; new erfBackward. The legacy RecordingTensorOpsDecorator now records gelu/erf instead of delegating them unrecorded (before this PR a compiled graph never saw GELU at all).
• StableHLO: erf → chlo.erf %x : T -> T (CHLO unary assembly); exact GELU → 0.5·x·(1 + chlo.erf(x·√½)). One GeluLowering builder emits both the tanh chain (unchanged text) and the exact chain through an HloEmitter seam, so part 2 can wrap the same builder into a stablehlo.composite decomposition without a second copy of the formula.
• DSLs / API: network {} gets gelu(approximate: Boolean = true, id) and erf(id) (the KSP default tables are name-keyed — approximate was added to both generators); GELU(name, approximate) module; Java TensorJavaOps.gelu(a, approximate) / erf(a), SequentialModelBuilder.gelu(approximate); ground-truth OperationParams.approximate.

Breaking

TensorOps.gelu gained a parameter: implementors of the interface must update the override (callers are source-compatible through the default). In the generated network DSL the activation convention puts parameters before id, so a positional gelu("name") becomes gelu(id = "name") (same as leakyRelu(slope, id)).

Test plan

• skainet-backend-cpu: erf reference table (2e-7), oddness/saturation, exact gelu bit-equal to the host formula, default byte-identical to the old tanh path, exact-vs-tanh band 4.5e-4..5.0e-4 at |x| ∈ [2.5, 2.9], Int32 rejection; JVM erf_and_exact_gelu_match_scalar_reference_with_tails
• skainet-compile-dag: GeluExactRecordingTest (flag on trace and node, DIRECT vs OPTIMIZED replay ≤ 1e-6), exact-gelu and erf finite-difference gradients, AutodiffCoverageTest (erf rule registered)
• skainet-compile-core: legacy recorder records GeluOperation(approximate) / ErfOperation
• skainet-compile-hlo: testGeluOperation unchanged, testGeluExactOperation, testErfOperationLowersToChlo, approximate=true equals default
• skainet-lang-core: generated DSL gelu(approximate = false, id = …) compiles and builds; KSP processor tests
• apiDump, generateDocs, scripts/pr-gate.sh --quick

Part 2 (composite emission behind OpGranularityPolicy, explicit exact→tanh legalization pass, @Lowering matrix column) follows in a separate PR, as laid out in #1342.