[ROCm][inductor] Test dynamic-shape TunableOp wildcard dispatch (#192144) (#192144)

Summary:

Test is split out due to large commit size.

Add unit coverage for wildcard persistence, runtime fallback, layout remapping, safe non-tunable fallback, and dynamic-dimension mask derivation.

Test Plan: HIP_VISIBLE_DEVICES=0 buck2 test fbcode//mode/dev-nosan-amd-gpu --modifier rocm70 -c fbcode.enable_gpu_sections=true -c fbcode.triton_backend=amd -c fbcode.rocm_arch=mi350 fbcode//caffe2/test/inductor:dynamic_tunable_ops

Reviewed By: Ruishenl, ZhitongGuo

Differential Revision: D114394201

Pull Request resolved: https://github.com/pytorch/pytorch/pull/192144
Approved by: https://github.com/mlazos