ImageNet benchmark runs can't cleanly resume: no last.ckpt + timestamped run dirs

Nobody has claimed this yet.

`--ckpt-path` already flows into `trainer.fit(ckpt_path=...)` (`benchmarks/imagenet/resnet50/main.py:323`, `benchmarks/imagenet/vitb16/main.py:269`), so Lightning restores epoch, optimizer and scheduler state. Two gaps stop it being a clean resume.

**1. No explicit `ModelCheckpoint`.** The `pretrain()` callback lists (`resnet50/main.py:306`, `vitb16/main.py:252`) rely on Lightning's implicit epoch-end checkpoint. No `last.ckpt`, no mid-epoch save, no retention control for 100-epoch runs.

**2. Every launch stamps a fresh run dir** with `datetime.now()` (`resnet50/main.py:122`, `vitb16/main.py:103`). A resume logs to a new folder and there is no discovery of the prior checkpoint, so you hand-locate the `.ckpt` path.

**Scope** (benchmark-only, both `main.py`):
- [ ] Add `ModelCheckpoint(save_last=True, ...)` to both `pretrain()` callback lists.
- [ ] Add `--resume` that resolves the newest run dir + its `last.ckpt` and reuses that dir instead of stamping a new one.
- [ ] Pass the resolved path to `ckpt_path`.
- [ ] Document the flag in `benchmarks/imagenet/README.md`.

Read benchmarks/imagenet/resnet50/main.py and vitb16/main.py around pretrain(), the callback lists, and the timestamped run-directory setup. Trace how --ckpt-path reaches trainer.fit, then implement and verify --resume uses the newest run directory and last.ckpt for both benchmarks. Update benchmarks/imagenet/README.md to document the flag and expected resume behavior.

Written by the indexing model from the issue text.