suspendmanager: suspend unsafe machine deconstruction

[Alistair-Afton]

Summary

Fixes #5777.

suspendmanager now inspects DestroyBuilding jobs: if removing a machine component would leave a connected component of its machine graph without any terrain-anchored node, the job is suspended with the new reason "Would collapse machinery". Deconstructions that leave every remaining component anchored proceed normally.

• New Reason::DANGER_OF_COLLAPSE plus a union-find pass over machine_nodest connections that checks each surviving connected component for an anchor
• A node counts as anchored only if its entire footprint sits on non-open tiles
• The unsuspend pass and the building-panel overlay now handle destroy jobs so the reason is visible in-game
• Scope note: only the machine containing the destroyed building is considered; components in other machines that are already orphaned collapse on the next stability pass regardless of which removal triggers it

Machine-collapse semantics (verified in-game on 53.16)

• world.buildings.check_machine_collapse triggers a stability pass over all machines after a building removal
• A connected graph component survives iff it retains at least one anchored node; otherwise it collapses wholesale
• isPowerSource() alone does not anchor: a lone waterwheel over open space collapses
• A single anchored gear holds up arbitrarily long chains of axles/gears hanging over open space
• Orphaned components can persist indefinitely until some removal triggers the next pass

Testing

• New test/plugins/suspendmanager.lua case fabricates a floor-anchored gear, hanging axle, and hanging gear chain, queues destroy jobs, runs suspendmanager now, and asserts the anchor's job is suspended while the endpoint's is not
• 7/7 tests, 21/21 checks pass in-game
• Also verified live on a purpose-built rig: the suspended anchor job prevented the collapse and the endpoint removal proceeded normally

[Alistair-Afton]

The gcc-11 leg failure is not from this change: DF's test suite reported all tests passing, then crashed at quit with free(): invalid pointer, and the same signature just failed an unrelated branch's run in the same window. Filed as #5947. Windows legs hit the known retry-timeout issue that #5943 fixes.