[bug] int32 signed overflow in buffer x

[darxradi3nt]

offset corrupts gradients silently for large batch/dstate configs

The buffer x pointer offset in both the forward and backward selective-scan CUDA kernels is computed using an all-int32 multiply.
When batch * dim * n_chunks * dstate exceeds max int32, it wraps the result negative.
The kernel then reads and writes memory located before x_ptr, silently corrupting adjacent tensors or triggering a CUDA illegal-address fault.

fixed in pr #883

[darxradi3nt]

@tridao @albertfgu I would be grateful for your kind feedback