Expose resolved GEMM schedules in library descriptions

[zupengwang]

The instance library currently omits the mainloop and epilogue scheduling policies from GEMM descriptions, so clients must parse kernel names to select between cooperative, ping-pong, 1-SM/2-SM, and FP8 fast-accumulation variants. Expose the resolved mainloop schedule, operand load kind, epilogue strategy, and accumulation kind as runtime enums. The mappings use collective dispatch policies and MMA CTA groups, including Auto schedules, and populate dense, grouped, sparse, blockwise, and block-scaled descriptions.

Existing constructor arguments remain compatible, and unrecognized policies retain explicit unknown values. The public description structures gain fields, so the instance library and clients must be rebuilt together. No kernel computation or scheduling implementation is changed.

Validation: the API probe fails on the exact base because the four fields are absent and passes with this change. The new host-executed unit target passes 19 tests with CUDA 13.0, 9 with CUDA 12.1, and 3 with CUDA 11.8; architecture cases are gated by toolkit support. CTest passes for all three builds. Tests ran with GPUs hidden, and a syscall trace observed no GPU device opens. This validates metadata and compilation, not Hopper/Blackwell GPU execution or performance.

Addresses #2073.

AI-assisted implementation, reviewed by the submitting author.