* llama : fix DSV4 compressor padding for simple batches

Count actual tokens per sequence when padding compressor writes. Add a graph reuse regression test for odd-sized DeepSeek V4 and V4.1 batches.

Assisted-by: Codex

* tests : cover multiple sequences in DSV4 padding regression

Exercise one and two sequences at different positions while keeping a fixed ubatch shape.

Assisted-by: Codex

* llama : remove simple batch splitting for DSV4

Use the existing sequence splitting to keep n_seq_tokens consistent with
the compressor padding calculation. Extend the graph reuse regression to
6-token batches.

Assisted-by: Codex