Replace np.random.randn/rand/seed with np.random.default_rng(7) across
all 93 source modules, tests, and milestones for reproducible, isolated
random state.
TinyDigits pkl files were stored in LFS, causing install failures
when users don't have git-lfs installed. The sparse checkout would
download LFS pointer files instead of actual data, causing pickle
to fail with 'invalid load key' errors.
These files are small (~300KB total) so LFS is unnecessary.
Fixes#1129
Add curated educational datasets for TinyTorch milestones:
TinyDigits (~310 KB):
- 1000 train + 200 test samples of 8x8 digit images
- Balanced: 100 samples per digit class (0-9)
- Used by Milestones 03 (MLP) and 04 (CNN)
- Created from sklearn digits, normalized to [0,1]
TinyTalks (~40 KB):
- 350 Q&A pairs across 5 difficulty levels
- Character-level conversational dataset
- Used by Milestone 05 (Transformer)
- Designed for fast training (3-5 min on laptop)
Both datasets follow Karpathy's ~1K samples philosophy:
- Small enough to ship with repo
- Large enough for meaningful learning
- Fast training with instant feedback
- Works offline, no downloads needed
