This fixes RWKV inference which otherwise fails when ubatch.n_seq_tokens is 0.

I noticed this when trying to run the command from #13834 (comment), but with RWKV instead of Mamba.

For example, with https://huggingface.co/latestissue/rwkv-6-finch-1b6-gguf/blob/main/rwkv-6-finch-1b6-Q4_0.gguf

Before:

$ ./bin/llama-parallel -m /path/to/rwkv-6-finch-1b6-Q4_0.gguf -np 5 -ns 8 --temp 0 --repeat-penalty 1.1 -ub 2 -pps
.../ggml/src/ggml.c:1593: GGML_ASSERT(view_src == NULL || data_size == 0 || data_size + view_offs <= ggml_nbytes(view_src)) failed

After:

$ ./bin/llama-parallel -m /path/to/rwkv-6-finch-1b6-Q4_0.gguf -np 5 -ns 8 --temp 0 --repeat-penalty 1.1 -ub 2 -pps
(proceeds normally)

This has been broken since #13746 because it made the worst case graph_reserve use n_seqs = cparams.n_seq_max instead of n_seqs = 1.

ubatch.n_seq_tokens must not be 0 for RWKV, because in some views it uses ubatch.n_seq_tokens - 1 for one of the dimensions:

There will hopefully eventually be automated tests which will help detect this kind of problem, see #14139 (still in the early stages, though).

Make sure to read the contributing guidelines before submitting a PR