Data Exchange, STEP - Make the non-manifold detection flag per-actor

[gsdali]

STEPControl_ActorRead.cxx carries NM_DETECTED as an anonymous-namespace global. It is reset at the start of each shape-representation transfer, set when a non-manifold item is recognised, and read by the product-definition overload to decide whether an assembly component's COMPOUND is flattened into its parent or kept nested.

Two concurrent STEP reads therefore share a flag that gates shape construction.

The declaration's own comment already records the intended direction:

Set global var to inform outer methods that current representation item is non-manifold.
The better way is to pass this information via binder or via TopoDS_Shape itself, however,
this is very specific info to do so...

The fix

A private member with a default member initialiser. Every read and write is already inside a STEPControl_ActorRead member function, so no signature change is needed.

This is sufficient because STEPControl_Controller::ActorRead() never assigns myAdaptorRead, so it constructs a fresh actor per call which XSControl_TransferReader::Actor() caches per session. Each concurrent read owns its flag.

Measured

ThreadSanitizer, six threads reading a non-manifold and a manifold assembly concurrently, ten iterations each:

	NM_DETECTED race reported	Total reports
Before	5 of 5 runs	40 across 5 runs
After	0 of 5 runs	10 across 5 runs

What is NOT claimed

A wrong shape was not reproduced. The flag demonstrably leaks between threads (25 to 211 leaked reads depending on configuration), but across roughly ten thousand reads in three configurations zero shapes came back altered, because within a thread the reset reliably wins the race to that same thread's read.

So this is a confirmed data race on a flag that gates shape construction, not a demonstrated wrong answer. The GTest reflects that honestly: it pins the ownership invariant the fix relies on, that ActorRead hands out a distinct actor per call, rather than asserting an outcome that cannot be reproduced deterministically. If you would prefer a concurrency test that asserts the shape instead, I can add one, but it would be timing-dependent and I would not want to put a flaky test in your CI.