Allow to force to use full precision when computing attention

[Warvito]

As mentioned in Stable Diffusion v 2.1 (https://github.com/Stability-AI/stablediffusion/blame/main/README.md#L15), half precision in the attention computation might cause some instabilities. One thing that we could implement in our models to avoid this is the option to perform the attention with full attention or not

[Warvito]

Reference https://github.com/Stability-AI/stablediffusion/blob/8bde0cf64f3735bb33d93bdb8e28120be45c479b/ldm/modules/attention.py#L175