llama : add position-relocatable KV range save/load

[Anyesh]

Overview

Adds llama_state_seq_get_range / llama_state_seq_set_range: serialize a contiguous KV range [p0, p1) of a sequence and restore it at a different position, re-anchoring K via the existing deferred RoPE shift (no forward pass on restore).

get_data / set_data only restore a sequence at its original positions; there is no way to relocate a sub-range. This enables cache schemes that drop spans and later splice them back at a new offset without recompute. Works for plain, iSWA, and hybrid attention caches; no change to existing formats or public APIs.

Additional information

range_read places the saved cells at their original positions, then applies pos_add(new_pos - original_pos) to queue the same K rotation seq_add uses. tests/test-state-seq-range.cpp (added to CTest) covers round-trip, partial restore, and positive/negative/offset relocation. Relocation cases assert the result is bit-identical to an in-place llama_memory_seq_add shift rather than a from-scratch decode: the K-shift is not bit-exact versus a fresh rope, so greedy tokens can differ on near-tied logits even when relocation is correct.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes, used an AI coding assistant.