Updated the SD UNet definition to support both v1.5 and v2.1. 

Tested with

```
import tvm
from diffusers import StableDiffusionPipeline, EulerDiscreteScheduler
import web_stable_diffusion.trace as trace
from tvm.relax.frontend import detach_params

# pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")

model_id = "stabilityai/stable-diffusion-2-1-base"
scheduler = EulerDiscreteScheduler.from_pretrained(model_id, subfolder="scheduler")
pipe = StableDiffusionPipeline.from_pretrained(model_id, scheduler=scheduler)

unet, params_unet = detach_params(trace.unet_latents_to_noise_pred(pipe, "mps"))

```

This PR fixes the database of workloads to include the purity field
which was introduced to TVM Unity not long ago.

"the the" => "the"

1. As i verifed, the latents scale is same for sd/sdxl, otherwise the
image quality is terrible.
2. Add attention to remind use the right vae net.
fix mlc-ai#60