HloGenerator.generate emits an empty module — it doesn't push the tape onto Execution.tapeStack

HloGenerator.generateBlocking(model, input, name) returns a structure-only module
(func.func @<name>() -> () { return }, no stablehlo.* ops) for any Model whose
forward records through the tensor-bound ops — because it never activates the tape on
Execution.tapeStack, so the recorded operations don't land on the tape and
toComputeGraph() is empty. (This is the root cause of the "structure-only export"
noted in #663.)

The working recipe (from SKaiNET-transformers GemmaMlirDumpTest)

The transformers repo traces real models correctly by pushing the current tape:

val ctx = DefaultGraphExecutionContext.tape(baseOps = VoidTensorOps())
val tape = ctx.record {
    val ct = (this as DefaultGraphExecutionContext).currentTape ?: error("no tape")
    Execution.tapeStack.pushTape(ct)                 // <-- HloGenerator is missing this
    try { model.forward(input, this as ExecutionContext) }
    finally { Execution.tapeStack.popTape() }
}.first
val graph = (tape as DefaultExecutionTape).toComputeGraph(synthesizeExternalInputs = true)
val mlir = toStableHlo(graph, "gemma").content       // real ops (rmsNorm/sdpa/rope/...)

With pushTape, the same Model that HloGenerator exports as empty produces a full graph.

Ask

HloGenerator.generate (and generateBlocking) should push/pop the recording tape on
Execution.tapeStack around the forward call (and use
toComputeGraph(synthesizeExternalInputs = true)), so the documented one-call API matches
the working manual recipe. A test that asserts a non-empty module for Rgb2GrayScaleMatMul
(or any tensor-ops model) would lock it in.

cc #663