operators: emit trace configuration when trace_size is set

[atassis]

PR A -- amd/IRON, branch feat/iron-trace-parity-devel @ 4d9f60c (base devel)

Title: operators: emit trace configuration when trace_size is set

---

Most designs in iron/operators accept a trace_size argument but never call
Runtime.enable_trace, so they run fine and return a trace.txt of all zeros. That reads as
"this op can't be traced" rather than "the wiring is missing", which is what sent me looking.

Affected: axpy, binary_elementwise, channeled_unary (so gelu/silu/relu/sigmoid/tanh),
gemm, leaky_relu, mem_copy, mha, rope, softmax.

Rather than repeat the same event setup per design, the wiring goes in
iron/operators/_trace.py. maybe_enable_trace() uses an explicit trace_size when one is
passed and otherwise falls back to IRON_TRACE_SIZE, so the designs whose op.py doesn't plumb
trace_size through can still be traced; IRON_TRACE_NTILES caps how many workers are
instrumented (0 traces none). It's a no-op when neither is set, so default behaviour is unchanged.

Three things I'd rather flag than have you find

The import is absolute, not relative, and that matters. DesignGenerator loads these files by
path via spec_from_file_location + exec_module with no package context
(iron/common/compilation/base.py), so a relative import raises "attempted relative import with no
known parent package" at generation time. It also passes py_compile, so it won't show up in a
lint pass. Worth knowing before anyone tidies it. This is also the first dependency from a
design.py onto the surrounding iron package -- if you'd rather designs stayed standalone, say
so and I'll inline the helper instead.

dequant is deliberately untouched. It already wires trace, but in a simpler form (no
workers, no coretile_events), so folding it onto the helper would change its behaviour rather
than just deduplicate. Happy to do it as a follow-up if you want one path.

No automated coverage for the fix itself. The verification below is a manual check; nothing in
CI would catch this regressing again. If you want a test asserting a design emits aie.trace when
trace_size is set, tell me the shape you'd prefer and I'll add it.

Testing

Generated softmax through that same package-less loader: 0 aie.trace ops before this change
with tracing requested, 18 after, and still 0 when tracing is off. The 18 are a full
aie.trace @trace_core_1 block (8 core events, 3 port configs, start/stop broadcasts,
host_config {buffer_size = 8192}).

On device (Strix, NPU2): 935 tests pass across gelu, softmax, axpy, mem_copy, rope.
black --check and reuse lint both clean.