Enable smem_merge_branch_allocs option on branched mega-kernel examples

[LongshengDu]

4.6.dev version align kernel shared memory allocator default behavior with CUDA C++ (i.e. all static smem allocation will be counted).

Since mega-kernel approach has 2 mutually exclusive code branches, only one path runs per launch, thus new launch option smem_merge_branch_allocs is introduced for mega-kernel to enable shared memory reuse between two paths.

OFF (default)          ON
┌────────────┐         ┌────────────┐ ┐
│ Branch A   │ 512B    │ Branch A   │ │
├────────────┤         ├────────────┤ ├─ reuse same region
│ Branch B   │ 512B    │ Branch B   │ │
└────────────┘         └────────────┘ ┘
Total = 1024B          Total = 512B