fix: remove double softplus on dt in non-fused Mamba2

[Mr-Neutr0n]

Bug

In the non-fused/reference path of Mamba2Simple, softplus is applied to dt twice — once explicitly before passing to the SSM step, and once inside the step function. This over-smooths the timestep values.

Fix

Removed the redundant softplus application to match the fused kernel behavior.

[albertfgu]

Sorry, can you describe the exact code path where the softplus is applied twice?

It seems you're referring to the path starting line 158? Are you claiming the softplus is applied on line 162 and again on line 182 (mamba_chunk_scan_combined)? But I don't think the latter applies the softplus by default.

[darxradi3nt]

mamba_chunk_scan_combined defaults to dt_softplus=False, so it receives the already-transformed dt and applies no further softplus. softplus is applied exactly once.
The proposed change moves the operation inside the Triton kernel via dt_softplus=True, which produces the same result. That may be a cleaner factoring, but there's no double application to fix in the current code.

Suggest re-labeling as a refactor/cleanup rather than a bug fix and close the PR

[albertfgu]

Yes, if you relabel this I can merge it

[darxradi3nt]

@Mr-Neutr0n?