feat: add max-vram based segmented param offload

[leejet]

• Added support for --max-vram to limit runtime GPU memory usage.
• When a VRAM budget is set, the runner executes in segments based on tiled results, temporarily offloading only the parameters needed for the current sub-graph into VRAM instead of loading all parameters into GPU memory at once.
• This allows larger models to run on devices with limited VRAM and keeps peak VRAM usage close to the specified budget; if not set or set to 0, the original full-graph execution behavior is preserved.