Allow external adjustment of parallel search workers

[zayenz]

A portfolio may need to move a fixed CPU budget between search engines without rebuilding them. This PR adds Search::WorkerControl, a copyable handle that changes how many pre-created workers an engine may execute.

The initial threads value remains the fixed resident capacity. request() is thread-safe and does not wait for the engine to reach the requested count. Growing wakes parked workers; shrinking lets an in-flight search action finish before its worker parks. Zero pauses search without discarding its state, and a positive request resumes it. DFS and BAB also honor pause requests when the resolved capacity is one, including requests made during next().

Controls apply to DFS and BAB leaves owned by RBS and PBS. PBS keeps one control per asset; allocation policy and the portfolio-wide budget remain with the caller. Explicit controlled assets require parallel outer PBS execution; sequential outer PBS rejects them with WorkerControlInUse because a paused asset would prevent it from running siblings. A paused asset in parallel PBS observes the internal stop and reports when another asset closes the round. Its request remains zero and its search state is preserved. If every active asset is paused, next() waits until one resumes. TimeStop has no timer thread to wake a paused engine.

A control binds to one logical leaf-engine lifetime. Copies share request state, while an empty handle preserves fixed-worker behavior. Builds without thread support reject an initially paused attachment and request(0) with InvalidWorkerRequest; controls in those builds are restricted to use from one thread.

PBS retains ownership of each asset seed while building the asset with clone=true. Nested construction failures can therefore clean up without deleting the same space twice. This adds one space clone per asset during construction; search execution is unchanged.

Validation:

• The complete 25,812-test Search suite passed during validation in a Debug build. After the final configuration checks were added, all 100 affected tests passed for three iterations each in both threaded and no-thread builds. The no-thread normal check target also passed.
• Focused regressions cover nested construction failures and owned-space cleanup, initial and mid-search capacity-one pauses, resumed DFS/BAB enumeration, and paused DFS/RBS portfolio assets at capacities one and two.
• Twelve concurrent resize trials cycling through zero to four workers each enumerated all 1,024 expected solutions.
• The new construction and worker-control regressions are included in CMake and Autoconf check selections.

The Windows vcpkg job uses one current commit for both its checkout and builtin baseline. The previous baseline selected a GMP portfile referencing an MSYS2 autoconf archive that is no longer available.