llama : add memory-mapped session cache support
Add support for memory-mapped session caches, both read-only and
read-write. This provides efficient access to session files through
memory mapping, which can reduce memory copies and improve performance
for large session files.
New features:
- llama_state_mmap_open() - Open session file for read-only mmap access
- llama_state_mmap_open_rw() - Open/create session file for read-write
- llama_state_mmap_load() - Load state from mmap into context
- llama_state_mmap_save() - Save state from context to mmap (rw only)
- llama_state_mmap_sync() - Sync changes to disk (rw maps)
- llama_state_mmap_free() - Free the mmap handle
- llama_state_mmap_supported() - Check platform support
Implementation details:
- Adds llama_mmap_rw class supporting both read-only and read-write mmap
- Uses POSIX mmap on Linux/macOS and CreateFileMapping on Windows
- Read-write maps automatically resize when saving larger states
- Changes are synced to disk on save and when freeing the handle
