Replace the network-wide init scheme with per-parameter initializers

[jessegrabowski]

initialize_params and Model.initialize take initializers={parameter: HeNormal()}, keyed by the parameter object rather than its name, which is the only way to reach something no constructor keyword exposes — a bias inside a transformer block, or a weight in a network you loaded rather than built.

Then scheme= goes. It was the one API that acted on parameters the caller hadn't named, and it was paying for itself with everything around it: Linear.W had to withhold its own default so a scheme could reach it, which made a recorded declaration a silent library-wide breakage; a constant had to be fenced off from it; precedence had to be three-way. All of that is gone, and Linear and Embedding now declare the Xavier draw they were already using. initialize(seed=0) means redraw every parameter from its own initializer.

model.initialize(seed=0, initializers={block.ff.fc_in.b: NormalInitializer(0.0, 0.02)})

One behavior change worth a look: a parameter that declares no initializer now raises rather than being silently left alone. Only reachable by building a trainable() by hand — every layer declares one per parameter — but leaving it alone would put an all-zero weight back in a graph that trains, which is what #84 just fixed.

Also worth knowing: he_normal across a whole network is now unsayable, and permanently, since once W declares there is no set of undeclared weights left to address. Per-layer weight_initializer= is the replacement, as in torch, flax and keras.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.48%. Comparing base (5c54d4a) to head (91952ca).

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #85      +/-   ##
==========================================
- Coverage   95.49%   95.48%   -0.01%     
==========================================
  Files          45       45              
  Lines        1819     1818       -1     
==========================================
- Hits         1737     1736       -1     
  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.