Allow disabling grouped GEMM for Mamba/hybrid

Originally posted here:
https://github.com/NVIDIA/Megatron-LM/pull/4099/changes#r3026195753
cc @kevalmorabia97

**Is your feature request related to a problem? Please describe.**
The Mamba/hybrid model currently does not support setting the `moe_grouped_gemm` via `config` in `megatron/core/models/mamba/mamba_layer_specs.py`/`megatron/core/models/hybrid/hybrid_layer_specs.py`, unlike the GPT model. ModelOpt pruning does not support grouped GEMM, so this would enable easier support without having to modify the code itself.

**Describe the solution you'd like**
(Refers to `megatron/core/models/mamba/mamba_layer_specs.py`, which may have been renamed to `megatron/core/models/hybrid/hybrid_layer_specs.py` in the meantime.)

> Can we add a convenience function `get_te_mamba_stack_spec(moe_grouped_gemm: bool = False)` so we can configure whether to use moe grouped gemm or not? This is similar to gpt model spec function - this is useful for ModelOpt pruning where we dont support grouped gemm