trunk/803b5d8d9e5b41e39c1d9a11537c30efdded9230: [Inductor][NVGEMM] Refactor rendering (#186184)

Replaces the Jinja template (nv_universal_gemm.py.jinja) with

IndentedBuffer codegen in NVUniversalGemmKernel.render(). Introduces

dataclasses (_VariantRenderSpec, _EpilogueRenderSpec) and module-level

helpers (_create_gemm_arguments, _create_gemm_cache_key,

_lookup_gemm_kernel) that are imported by the generated wrapper at

runtime.

Restores disk cache (disk_cache_get/set) and the _precompile hook for

subprocess parallel compilation that were present in the Jinja template.

Cache keys now include strides, output metadata, and device for all

tensor variants.

Authored with the help of an AI assistant (Claude).

Test Plan:

python test/inductor/test_nv_universal_gemm.py -v

Pull Request resolved: #186184

Approved by: https://github.com/drisspg

ghstack dependencies: #186183