server: detect conversation boundaries via conv_hash to fix session bleed + K-shift NULL buffer crash

Three fixes:

1. Conversation boundary detection (fixes bugs 1+2: match nosedives
   after harness trims, new sessions reusing stale KV cache):
   - Preserve conv_hash across reset() instead of zeroing it, so
     launch_slot_with_task can compare the previous session's hash
     with the incoming task's hash
   - In get_available_slot: compute task conv_hash, skip LCP-matched
     slots with mismatched conv_hash, clear stale KV cache + prompt
     tokens before prompt_load so the cache search targets the new
     session
   - In launch_slot_with_task: safety-net conv_hash check + clear

2. K-shift crash on ISWA caches with disparate sizes (Laguna XS.2
   with kv-unified): when build_graph_shift creates k_shift/k_rot
   input tensors but no layers have RoPE to reference them, the
   scheduler never allocates a buffer. set_input_k_shift then
   crashes on GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer)).
   Fix: early-return if no layers have RoPE, plus a defensive
   buffer-null guard in set_input matching the pattern already
   used by llm_graph_input_attn_kv and llm_graph_input_dsv4_raw.