[Fix] deepcache uncommon resolution base

[lixiang007666]

This PR is done:

• Fix Diffusers image-to-image with odd dimensions are broken #602 .

Test:

import torch
from onediffx import compile_pipe
from PIL import Image
from onediffx.deep_cache import StableDiffusionXLPipeline

t2i_pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    variant="fp16",
)

t2i_pipe.to("cuda")
t2i_pipe = compile_pipe(t2i_pipe)

from diffusers import AutoPipelineForImage2Image

i2i_pipe = AutoPipelineForImage2Image.from_pipe(t2i_pipe)
i2i_pipe.to("cuda")

# Works
for dimensions in [(512, 512), (544, 544), (576, 578)]:
    i2i_pipe(
        prompt="a photo of a cat",
        negative_prompt="",
        image=Image.new("RGB", dimensions),
        num_inference_steps=20,
    )

# Doesn't work
i2i_pipe(
    prompt="a photo of a cat",
    negative_prompt="",
    image=Image.new("RGB", (528, 528)),
    num_inference_steps=20,
)