Add solution articles for multi-layer backpropagation, weight
initialization, and batch normalization. These follow the same
format as the existing 27 ML solution articles (Prerequisites,
Concept, Solution with Python tabs, Common Pitfalls, In the GPT
Project, Key Takeaways).

Made-with: Cursor

- multi-headed-self-attention: Add output_proj (W_o) linear layer after
  concatenating heads, matching standard practice
- transformer-block: Add output_proj to inner MultiHeadedSelfAttention
  class, consistent with multi-head attention problem
- weight-initialization: Rewrite check_activations to use raw weight
  matrices (torch.randn * std) instead of nn.Linear for cross-platform
  determinism

Made-with: Cursor

Reduces precision from 4 to 2 decimal places for the
check_activations method to absorb cross-platform floating
point differences in multi-layer matrix operations.

Made-with: Cursor

- Remove softmax from solution, return raw logits
- Add W_o output projection to inner MHA class
- Update all explanatory text: probabilities → logits
- Update shape table and key takeaways

Made-with: Cursor

- rms-normalization.md — RMSNorm vs LayerNorm, numpy implementation, Llama/Mistral context
- training-diagnostics.md — 3-method class (activation stats, gradient stats, diagnose), Karpathy debugging recipe

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Two new ML course problems (PR4), bringing total to 34.

Dead ReLU Detector: detect_dead_neurons + suggest_fix with priority-ordered
diagnostic rules. 8 test cases covering all fix branches.

Tokenization Edge Cases: greedy tokenization exposing number inconsistency,
token counting, and fertility scoring. 8 test cases.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Made-with: Cursor