feat(hooks): add pluggable hook modes

[galletas1712]

Closes #77 (pluggable memory-saver implementations) — the torch_memory_saver piece of the GPU Memory Service integration proposed in sgl-project/sglang#27310.

• register_hook_mode(name, factory, uses_preload=...) registers a lazily constructed custom implementation, selected through the existing hook_mode setter. Factories are process-global and resolved on first use; the builtin preload/torch modes are registered the same way and behave exactly as before.
• The package-level configure_subprocess() now consults the selected mode's uses_preload flag and skips the LD_PRELOAD setup for modes that do not use the preload hook (both builtins keep it, preserving current behavior).
• Selecting an unknown mode, re-registering a name, or using a mode never registered in the current process (e.g. a subprocess that skipped registration) fails with a descriptive error.

[gemini-code-assist]

Code Review

This pull request introduces a registry for hook modes (register_hook_mode) allowing custom lazy factories to be registered, and updates TorchMemorySaver to resolve its implementation lazily. The feedback suggests several improvements: implementing __getstate__ and __setstate__ on TorchMemorySaver to prevent pickling errors after initialization, simplifying the README example imports for consistency, and raising descriptive RuntimeErrors instead of cryptic KeyErrors in _ensure_initialized and _uses_preload if the selected hook mode is not registered in the current process.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[fzyzcjy]

(replied in #77)