First Block Caching Infra for diffusers

[quic-amitraj]

Wan-AI/Wan2.2-T2V-A14B-Diffusers

1. Added support of Non-unified Wan transformers ( as for 720p, unified approach was failing). This could be enabled using flag use_unified=False
2. Added support of first-block caching for non-unified approach, which can be enabled using flag enable_first_block_cache. Caching is only enabled with non-unified approach.
3. Added test for both non-unified and caching

Example-

pipeline = QEffWanPipeline.from_pretrained(
    "Wan-AI/Wan2.2-T2V-A14B-Diffusers",
    use_unified=False,
    enable_first_block_cache=True,
    # Hidden-dimension downsampling used for first-block residual similarity check.
    first_block_cache_downsample_factor=4,
)

Result- 90p/16TS/81f

black-forest-labs/FLUX.1-schnell

1. Enabled first block caching for flux
2. Caching can be enabled with flag enable_first_block_cache
3. Example-

pipeline = QEffFluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-schnell",
    enable_first_block_cache=True,
    # Hidden-dimension downsampling used for first-block residual similarity check.
    first_block_cache_downsample_factor=4,
)

4. Result-
   height/width=256