Refactor prompt parser

[AI-Casanova]

Full refactor of prompt parsing.

Moved prompt embeds inside of a class object that parses prompts at initialization and returns them when called.

Full scheduling capabilities restored

LRU cache for prompts, size adjustable from options

[vladmandic]

closes #2712

[vladmandic]

btw, i haven't commented on this for past few days because it looks ok, need to actually test it and i've been delaying that.

[vladmandic]

@AI-Casanova 2 critical bugs in the first 2min of testing...

scenario: sdxl, simple prompt (one sentence), no batching, no scheduling

breaks with cfg 0-1 (no guidance)

RuntimeError: The size of tensor a (2048) must match the size of tensor b (4096) at non-singleton dimension 1

then i noticed that all images look washed out.
in general, there is something really off with guidance - see two grids with NEW vs OLD code:

infotext for reproduction:

photo of a cute cyan robot with glowing glyph on a chest plate next to a road sign, next to a spaceship, walking sand with rocks and mountains and moon and planets visible in background, rich details, 4k
Steps: 15, Seed: 922182994, Sampler: UniPC, CFG scale: 2.0, Size: 1024x1024, Parser: Full, Backend: Diffusers, App: SD.Next, Version: e7f74c1, Operations: txt2img, Model: TempestV0.1-Artistic, Model hash: 9ca260b31e, Grid: 1x5x1, Pipeline: StableDiffusionXLPipeline, Script: XYZ Grid, X Type: [Param] Guidance scale, X Values: "2-10:5"

[AI-Casanova]

I have a sneaking suspicion that disabling the callback code will solve both issues.