v0.9.0

Added tinygrad backend and completely restructured use of neural networks

• refactor: drop deepmriprep from atlas and device code
• refactor: use t1prep.stages in probe and realign
• refactor: drop deepmriprep/deepbet from segment.py
• feat: add --nn-backend option to select network backend
• feat: add framework-neutral nn package
• feat: add stages package replacing deepmriprep
• fix: keep the sphere when fold removal fails
• fix: remove long-mode temporary warp fields
• fix: let deterministic mode run on CUDA
• fix: give --long-data one meaning, a file
• fix: remove long-mode p maps and average
• fix: verify downloads with requests, use interpolate
• fix: keep model weights in a release-named cache
• fix: pad progress bar lines so steps overwrite
• fix: fit progress frames to terminal width
• fix: follow phases and bars in web UI job page
• update Template_05mm_bet.nii.gz
• updated and added evaluation results
• perf: avoid slow deterministic interpolate path