Fix Mamba-3 decode path: forward() to step() shape handling

[sanowl]

Mamba3.forward() routes to step() during autoregressive decode (seqlen_offset > 0),
passing the (batch, 1, dim) tensor it receives. step() expects (batch, dim):
_preprocess uses 2-axis einops patterns such as rearrange(x, "b (h p) -> b h p").
Decoding through the module crashes at the first generated token. Even without the
crash, the previous return out returns 2D while the prefill path returns 3D.
Mamba2.step() handles this by squeezing the input and unsqueezing the output;
Mamba3 did neither.

the Fix

• Squeeze the singleton seqlen before step(), restore it after, with an
  assert seqlen == 1 guard (single-token decode, matching Mamba2).
• Add test_forward_decode_dispatch_matches_step: prefill via forward, decode
  through the seqlen_offset > 0 branch, and check parity against the direct-step
  reference (SISO and MIMO, out-norm on and off).

I have not run this. No NVIDIA GPU available locally (the CUTE step kernel requires
H100). The change is verified by code review and matches the existing Mamba-2
convention; the added test should validate it on CUDA / CI.