[ROCm][inductor] Dynamic-shape TunableOp: persist wildcard GEMM entries at autotune and dispatch them at runtime across all variants (mm/addmm/bmm/baddbmm/_scaled_mm) (#191658) (#191658)

Summary:

Persist wildcard GEMM tuning entries for dynamic shapes and dispatch them safely across TunableOp GEMM variants.

Test Plan: See the stacked unit-test diff.

Reviewed By: mlazos, ZhitongGuo, Ruishenl

Differential Revision: D105588444

Pull Request resolved: https://github.com/pytorch/pytorch/pull/191658
Approved by: https://github.com/mlazos