Add optimized inference scripts, memory utilities, and Studio documentation

[MrEdwards007]

This commit introduces the core optimized scripts and documentation for the TurboDiffusion framework:

• i2v.py: Optimized Image-to-Video inference featuring a Tiered Failover System and tiled encoding/decoding for 720p stability.
• t2v.py: Text-to-Video inference with intelligent hardware detection (TF32/BF16) and a 3-tier memory recovery system.
• cache_t5.py: A memory utility to pre-compute T5 embeddings, allowing 14B models to run without loading the 11GB text encoder into VRAM.
  TurboDiffusionStudio: (turbo_diffusion_t5_cache_optimize_v6.py) A unified Gradio-based Web interface for both Text-to-Video and Image-to-Video generation
• TurboDiffusion_Studio.md: Comprehensive documentation for the unified Gradio Web UI, detailing VRAM monitoring, frame sanitization, and auto-caching features.

All scripts include optimizations co-developed by Waverly Edwards and Google Gemini, and received support (cache_t5.py) by Johhn D. Pope, to maximize performance on consumer GPUs.

I was able to run this on a 5090, maxing out at 117 frames in 136 seconds (I2V) and 160 seconds (T2V), 16:9 aspect ratio using sagesla attention. Technically the T2V can be done in 126 seconds, using sla attention but I cant explain why sagesla seems to always outperform sla, except in this instance.

To reproduce without the gradio interface, utilize the committed files and this attached metadata, which can be used as a script.

t2v_20260101_085350_metadata.txt
i2v_20260101_063848_metadata.txt

(This is the original image used for I2V)

There are three anomalies.

1. Using the same seed does not always produce the same image. If I use a consistent seed, requesting 33 frames, the image may be different at 113 frames.
2. The I2V workflow has a subtle artifact that comes from the tiling and stitching them back together that I have NOT been able to eliminate. There is a tiny discontinuity/offset of the frames. I tried to perfectly align and smooth them out but it is still there.
3. When the video begins, it flashes or starts out brighter and then settles to the normal exposure.

(This is a frame of the video using the original image. Note the area [inside] of the red rectangle . When you look closely, there is a small gap and shift of a few pixels, that I've been unable to eliminate)