Draw every parameter at construction, and refuse a constant scheme

[jessegrabowski]

A network built and never .initialize()d couldn't learn. Every weight matrix was created as zeros, so nothing propagated below the last layer and only the output bias moved — the loss still crept down, which reads as slow convergence rather than a bug. Parameters are now drawn from their own initializer when the layer builds them, which incidentally makes Linear("fc", 4, 8, bias_initializer=NormalInitializer()) give you that bias instead of silently keeping zeros until someone calls initialize().

Second commit closes the other route to the same dead network: initialize("zeros") now raises. A constant is the right initializer for one bias or norm scale and never for every weight matrix in a network, so it stays available as a declaration on the parameter that wants it.

Construction draws from fresh entropy, so initialize(seed=...) is still the only thing needed for reproducibility and still overwrites everything.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.49%. Comparing base (620fbd1) to head (6748496).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #84      +/-   ##
==========================================
+ Coverage   95.47%   95.49%   +0.01%     
==========================================
  Files          45       45              
  Lines        1811     1819       +8     
==========================================
+ Hits         1729     1737       +8     
  Misses         82       82              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.