fix compatability issues with python 3.9

vae: fix attention blocks and timestep conditioning

VAE: Check for timesteps parameter in decoder before calling

VAE: Add timestep conditioning

Merge pull request #34 from LightricksResearch/add_atten_to_decoder

Decoder: add AttentionResBlocks block

Merge pull request #33 from LightricksResearch/compress-all-half-channels

causal_video_autoencoder: add option to half channels in depth to space upsample block

Merge pull request #29 from LightricksResearch/scale_vae_decoder_v1

VAE Decoder: Inject noise between conv layers.

Merge pull request #30 from LightricksResearch/fix-no-flash-attention

model: fix flash attention enabling - do not check device type at this point (can be CPU)

Feature: Add full bfloat16 support.

Merge pull request #19 from LightricksResearch/bfloat16-inference

Feature: Add mixed precision support and direct bfloat16 support.

VAE: Support different latent_var_log options when returning intermediate features for 3D perceptual loss

VAE: Support retuning intermediate features for 3d perceptual loss

Remove the word "pixart" from code.

VAE: Support more configurations for Encoder and Decoder blocks

Lint: added ruff.

VAE encode bugfix: remove bad dist_util.

CausalVideoAutoencoder: made neater load_ckpt.

Ckpt conversion: script + usage examples updated.

Merge pull request #7 from LightricksResearch/feature/fix-transformer-init-bug

transformer3d: init mode xora never happens because lower case needed.

import fixes.

Added tpu flash attention.

working.

refactor