v0.7.4

Changed

• sh/train_*.sh run detached and log everything. Every knob the scripts use is
  now one line in a run-settings block at the top (NAME="${NAME:-default}": DATA,
  OUTDIR, CFG, GPUS, BATCH_GPU, KEEP_LAST, AUGMENT, NUM_FID_SAMPLES,
  SEED, WORKERS, INIT_WEIGHTS, LR_WARMUP (512), CONDA_ENV,
  CUDA_VISIBLE_DEVICES, OMP_NUM_THREADS, MKL_NUM_THREADS, NCCL_DEBUG,
  HF_HOME, TORCH_HOME); the environment still overrides each one, and the
  resulting diffit-train command and exported environment are identical to 0.7.3's.
  On a workstation bash sh/train_<r>.sh re-launches itself with setsid nohup in its
  own session and returns, printing the log, tail -f and kill -- -<pid> lines; the
  output goes to logs/diffit-train_<r>-<date>.log with a .pid file beside it.
  FOREGROUND=1 stays attached and tees into the same log; under SLURM the script never
  detaches and tees into both the slurm .out and the log. Each log starts with a
  Run settings: block: every variable, the git commit (-dirty if the tree has
  changes), host, date and the full command line. LOG_DIR moves the logs.

Fixed

• README training advice matches the presets, the scripts and the paper. The
  "Strategy B" fine-tune recipe (written before the presets were settled) recommended
  --lr 5e-5 --lr-warmup 500 --kimg 100000 at 512² and --lr 2e-5 --lr-warmup 500 --kimg 50000 at 1024², and the "512² from scratch" example passed --lr-warmup 1000
  as "recommended"; none of that is what sh/train_*.sh run or what the paper uses.
  The section is replaced by one stage table (preset, LR, warmup, global batch, start)
  with the paper's App. I.2 wording quoted: LR 3e-4 / 1e-4 at every stage (the paper
  gives no fine-tune LR, training length, optimizer or warmup, and NVlabs/DiffiT ships
  no training configs), the 1000-kimg warmup only for warm-started stages (this fork's
  choice), from-scratch 512² without warmup. The unsourced "3–5× cheaper" claim, the
  walltime guesses and the 1024² --grad-accum 4 example (the preset uses 2) are gone.
  The README also notes the CFG provenance (official sample.py: 4.4 / 1.49; the
  paper text: 4.6 at 256). No preset or script hyperparameter changes.