feat: add lazy load eager eviction

[shikaku2]

Split out from draft PR #1573: #1573

Summary

Adds a -ll / --lazy-load flag for lazy load/eager evict model execution. The goal is to make larger image models practical on limited VRAM systems by loading model tensors through mmap and dropping already-used component pages after each pipeline stage.

This PR intentionally contains only the runtime lazy-load/eager-evict path. The conversion/RMSE/AIO work from #1573 is not included here.

How it works

• Enables mmap-backed model loading when --lazy-load is set.
• Forces free_params_immediately, so component backend buffers are released after use.
• Adds a platform-layer evict_memory() helper in util.cpp.
• Calls eager eviction after text encoder, diffusion, and VAE stages.
• Gates MADV_DONTNEED behind SD_MMAP_FLAGS=dontneed; without that flag, the helper is a no-op.
• Applies eviction to normal mmap-to-VRAM code paths as well, not just the new flag path.
• Auto-enables VAE tiling with --lazy-load to avoid very large single VAE decode allocations. On my Vulkan setup, SD3.5 VAE decode at 1024x1024 would otherwise hit a large VkBuffer allocation that exceeds common Vulkan maxMemoryAllocationSize limits.

Testing

• Built successfully with cmake --build build -j16.
• git diff --check passes.
• Runtime testing so far is on Vulkan only, on my machine (Arch Linux). That is the main reason this is a draft: it needs testing, and possibly follow-up code, on other GPU backends and operating systems.

Notes

This addresses the lazy-loading-specific review feedback from #1573 by keeping platform-dependent eviction code in util.cpp, making MADV_DONTNEED opt-in through SD_MMAP_FLAGS=dontneed, and calling eviction for regular mmap-backed load paths too.