Fix AMD preload startup and exit-time capture

[CodingInAVan]

HIP launcher capture previously stayed empty by default, and enabling the preload constructor crashed while assigning unconstructed C++ options. AMD-only Linux/glibc builds now initialize after shared-library constructors and before executable constructors, so early HIP code-object registration is visible. CUDA startup remains on its existing path.

Exit-time scope and telemetry access now respect thread-local destruction, disassembler helpers do not inherit the injection sentinel, and static inventory accepts the discrete Radeon reported as AMD Radeon Graphics.

Validation: CPU ABI/order regression (also added to CI), full AMD library preload tests for normal return and explicit exit with both constructor settings, 58 focused client tests, and allocation diagnostics for the original exit-time use-after-free. The content-pinned ROCm 7.2 candidate passed three contained RX 9070 XT trace fixtures (default, constructor opt-in, repeated default): correct 1,048,589-element result, AMD/gfx1201 inventory, one timed kernel, three full-size timed copies, captured source, 13 mapped ISA rows, clean shutdown, and confirmed container cleanup. Backend qualification evidence is committed in gpu-flight/gpufl-backend on amd-runtime-injection-fix. No other client PR is open.