selective scan cuda kernel x buffer overflow fix

[darxradi3nt]

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

scenarios that overflow:

Model	batch	dim	dstate	seqlen	n_chunks	batchdimn_chunks*dstate	First overflowing batch_id
Mamba-13B	32	5120	256	65536	32	1,342,177,280	ok
Mamba-7B	64	4096	256	67584	33	2,248,146,944	62
Mamba-13B	64	5120	256	65536	32	2,684,354,560	42

For Mamba-7B with batch=64 and seqlen=67584: items batch_id=62 and batch_id=63
write to invalid memory, producing wrong gradients silently on every training step
(~3% of the batch)