Fix SSM dt clamp default for Nemotron-H

[kernelpool]

Fix that removes incorrect lower bound clamping on SSM time steps for Nemotron-H, which negatively affects output quality. This also matches the HF reference. Results from tests below (4-bit versions):

seq_len=4096, 25 samples, tulu-3-sft-mixture

Model	PPL (before)	PPL (after)	Delta	KL div
Nano 4B	4.864	4.845	-0.019	0.006
Nano 30B	3.987	3.946	-0.041	0.034
Super 120B	3.439	3.363	-0.076	0.046

seq_len=4096, 25 samples, calibration_v5.txt

Model	PPL (before)	PPL (after)	Delta
Nano 4B	9.983	9.886	-0.097
Nano 30B	6.673	6.513	-0.159
Super 120B	5.563	5.203	-0.361

[Thump604]

Nice find — this matches the HF reference and the perplexity improvements are significant, especially on the Super 120B.

I run Nemotron-3-Super-120B-A12B (5-bit) daily on M2 Ultra 128GB via the Nemotron-H implementation from PR #992. Will test this fix against my production workload and report back with before/after comparison.

The original time_step_min → time_step_limit mapping was likely cargo-culted from Mamba-1 where it made sense; Nemotron-H's Mamba-2 doesn't need it. +1 for merge.

[angeloskath]

Awesome! Thanks for the fix and the evals.