memory : add GPU-to-GPU checkpoint save/restore for recurrent state

[petter-b]

cont #1 related upstream ggml-org#19493

Checkpoint-based speculative decoding serializes ~50 MiB of recurrent
state to host memory on every speculative cycle. On CUDA this costs
169-598 ms per save operation — 37% of total generation time.

This adds a device-to-device copy path using ggml_backend_tensor_copy,
allocating shadow tensors alongside the recurrent r/s tensors.

• llama_memory_checkpoint_save/restore/delete/supported public C API
  on llama_memory_i, implemented for llama_memory_recurrent
• Shadow tensor allocation in llama_memory_recurrent (~50 MiB lazily)
• Server callbacks wired to GPU copy path with serialization fallback
• Checkpoint timing instrumentation (save, restore, seq_rm durations)

Metal uses the same API but falls back to malloc→get→set→free on unified
memory (~0.4 ms per op — still negligible).

Per-operation checkpoint latency (RTX 3090, Qwen3.5-27B)

Path	Save latency	Restore latency
Serialization (host)	169-598 ms	~23 ms
GPU copy (device)	<1 ms	<1 ms

GPU copy eliminates the serialization bottleneck — per-operation cost
drops by 100-1000x.

Note: sustained throughput with ngram speculation is still limited by
draft acceptance rate, not checkpoint latency. See
petter-b/llama.cpp#1
for detailed throughput observations.

Tests

Existing checkpoint tests pass (10/10). New test:
test_spec_checkpoint_gpu_copy.py — verifies save/restore round-trip
correctness and timing.

Files changed (12 files, +530/-28)

• include/llama.h — public C API
• src/llama-memory.h — virtual checkpoint methods on llama_memory_i
• src/llama-memory-recurrent.h/cpp — shadow tensors, save/restore impl
• src/llama-context.cpp — API bridge
• tools/server/server-context.cpp — callback wiring, timing instrumentation
• 5 test files updated/added

Disclosure

Claude Code was used extensively for research, debugging, test generation, code implementation and writing this PR body. Validation is through TDD — each feature has a failing test before the implementation, verified on CUDA RTX 3090 and Mac-mini M4 Metal.

cc @srogmann