[CuTeDSL] Add BF16 grouped GEMM example for Hopper SM90

[vruga]

Closes #3040

Adds examples/python/CuTeDSL/hopper/grouped_gemm.py — a Python/CuTeDSL
grouped GEMM kernel for Hopper SM90 with BF16 (and Float16) support.

This is the CuTeDSL equivalent of examples/57_hopper_grouped_gemm,
which only exists in C++ and only supports FP8.

Design

• WGMMA + TMA with warp specialization (DMA warp group + MMA warp groups)
• Register accumulators (SM90 has no TMEM)
• PipelineTmaAsync for the A/B mainloop
• TensorMapManager for per-group TMA descriptor patching (SMEM and GMEM modes)
• StaticPersistentGroupTileScheduler for persistent multi-group scheduling
• Supports BF16/Float16 inputs, Float16/BFloat16/Float32 outputs

Tested on H100 (SM90) across BF16/Float16 dtypes, SMEM/GMEM tensormap modes, cluster shapes (1,1) and (2,1), up to 8 groups with unequal sizes.