mamba selective_scan CUDA kernel bugs

[darxradi3nt]

At the end of 2025 various code analysis engines were utilized to identify the following issues, including AI-based large language models. The same analysis done on the updated master branch at commit c1fcec8.

• smem_running_prefix write missing row offset
  read with a per-row offset but written without

• uint32_t stride overflow
  when overflow occurs, multiple batch_id values compute the same memory offset.
  Those batch items silently read/write the same memory, producing wrong outputs and wrong gradients.
  No CUDA error is raised. Realistic trigger: bs=17,dim=16384, seqlen=16384

more triggers:

Config	u_batch_stride	First bad batch_id	Wraps to
dim=16384, seqlen=16384	268,435,456	16 (bs ≥ 17)	0
dim=16384, seqlen=32768	536,870,912	8 (bs ≥ 9)	0
dim=16384, seqlen=65536	1,073,741,824	4 (bs ≥ 5)	0

• int32_t buffer overflow
  int32 signed overflow in x buffer offset corrupts gradients silently for large batch/dstate configs

PR: #881 #882 #883

[tridao]

sounds good, can you open a PR?

[darxradi3nt]

done, waiting for your review

[darxradi3nt]

#882 done with more casts, waiting for your review