[Example] Fix TileKernels Benchmarking

[WilliamZhang20]

List of new kernels:

• Per-channel cast
• Top-k
• mHC merge

New benchmarks:

Example	Shape	Tilus (ms)	TileKernels (ms)
Top-k	128 tokens, 72 experts, k=6	0.0065	0.0065
Top-k	1024 tokens, 256 experts, k=8	0.0084	0.0079
Top-k	8192 tokens, 256 experts, k=8	0.0178	0.0154
Per-token cast	128 × 1024	0.005824	0.006304
Per-token cast	256 × 2048	0.006208	0.006656
Per-token cast	257 × 4096	0.007008	0.007328
Per-token cast	8192 × 8192	0.073440	0.074688
SwiGLU + cast	128 × 1024	0.006080	0.030112
SwiGLU + cast	256 × 2048	0.007168	0.007552
SwiGLU + cast	257 × 4096	0.008288	0.009088
SwiGLU + cast	1024 × 4096	0.014304	0.015040
SwiGLU + cast	4096 × 8192	0.063392	0.064128
Per-channel cast	8192 × 8192	0.0797	0.0760
mHC merge	24 × 4096	0.0052	0.0052

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.