Adds epilogue fusion (relu, sigmoid, tanh, exp, add, sub, mul, div,
dtype cast) to the NVGEMM backend via cutlass_api EpilogueArguments.
The scheduler's fusion loop recognizes NVUniversalGemmCaller, benchmarks
fused vs unfused EFC kernels, and selects the winner.
MultiTemplateBuffer gains swap/finalize_as_nvgemm_caller for the
benchmark swap. When...