Use dynamic FLUX T5 seq len (small speed / mem improvement)

[RyanJDick]

Summary

This PR adds the option to dynamically select a T5 sequence length based on the length of the input prompt.

Using a smaller sequence length can improve inference speed and reduce peak memory, but will produce slightly different outputs.

Speed Improvement

Config: 1024x1024, bf16

T5 seq len = 512: 0.481 secs / it
T5 seq len = 128: 0.461 secs / it (4.1% speedup)

Memory improvement

Config: 1024x1024, bf16

T5 seq len = 512: 23355.16 MB peak VRAM
T5 seq len = 128: 23292.12 MB peak VRAM (63 MB saved)

Image difference

TODO: Add sample images showing the difference in output as T5 sequence len is varied.

Related Issues / Discussions

QA Instructions

• Test varied prompt lengths:
  • ~50 (t5 seq len of 128)
  • 127 (t5 seq len of 128)
  • 128 (t5 seq len of 128)
  • 129 (t5 seq len of 256)
  • ~250 (t5 seq len of 256)
  • ~500 (t5 seq len of 512)
• Test with regional prompts of varied lengths
• Test compatibility with:
  • ControlNet
  • IP-Adapter
  • LoRA

Merge Plan

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)