New comfy node for FlashVSR (supports int8, mitigates OOM)

I attempted to implement the LCSA logic as faithfully as possible, and also worked on compatibility with ComfyUI Dynamic VRAM and dedicated int8 convrot format.

This allows me to upscale to 1080p reliably on an RTX 4050 6GB VRAM 16GB RAM laptop where otherwise would OOM with other nodes.

This also supports Sparge Attention and the modified comfy kitchen library in https://github.com/Zironic/H3-Optimizations

I do notice that on fine details there seems to be seed dependent temporal inconsistencies. For example, an edge which is clear on the LQ video will have poor stability on the upscaled video, an issue that doesn't really appear when upscaling single images. Not sure if this is impact of the data I'm testing it on (low res MiniMax H3 clips using Turbo LoRAs) or something else; testing with online FlashVSR providers seems to result in similar issues.

My repo is here https://github.com/Pizzawookiee/ComfyUI-FlashVSR-Stock

My int8 convrot checkpoint is on Hugging Face: https://huggingface.co/pizzawookiee/FlashVSR-1.1/tree/main

Start by reviewing the external ComfyUI-FlashVSR-Stock repository and the linked Hugging Face checkpoint to understand the existing node and int8 format. Clarify whether the goal is upstream integration, and define the target files, tests, compatibility requirements, and acceptance criteria before starting.