Add missing template disambiguators

[Yinwei-Zhang]

Summary

Add missing template disambiguators to four dependent member-template calls in SM100 and SM103 GEMM kernels.

Root cause

CollectiveMainloop and CollectiveEpilogue are aliases of class template parameters, so member-template calls through their instances are dependent names. Standard C++ requires the template disambiguator in these expressions.

NVCC accepts the existing non-empty calls through permissive parsing, but stricter C++ frontends can reject them because < is not unambiguously parsed as the beginning of a template argument list.

Changes

• Disambiguate CollectiveMainloop::load_init and CollectiveMainloop::tensormaps_init calls in the SM100 array/grouped kernel.
• Disambiguate CollectiveEpilogue::load calls in the SM103 block-scaled GEMM and array/grouped GEMM kernels.

This is a syntax-only portability fix with no intended runtime or code-generation change.

Validation

• git diff --check
• Compiled 75_blackwell_grouped_gemm for sm_100a with CUDA 13.2.
• Built CMake target 89_sm103_fp4_ultra_gemm with CUTLASS_NVCC_ARCHS=103a.
• Built CMake target 90_sm103_fp4_ultra_grouped_gemm with CUTLASS_NVCC_ARCHS=103a.