Let a transformer block initialize its residual projections

[jessegrabowski]

Nothing could reach the projections inside a transformer block, so GPT-style init — scale the two that write back into the residual stream, leave their siblings alone — was inexpressible. One keyword per level routes it now: out_proj_initializer on attention, residual_initializer on the block, which feeds both that and the feed-forward's second layer.

Biases stay unreachable on purpose: pinned to zero by declaration, which is what the models this serves want, and bias=False covers the ones that drop them. A one-off gets the per-parameter override coming next.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.47%. Comparing base (f0b7651) to head (0b231c4).

Additional details and impacted files

@@           Coverage Diff           @@
##             main      #83   +/-   ##
=======================================
  Coverage   95.46%   95.47%           
=======================================
  Files          45       45           
  Lines        1809     1811    +2     
=======================================
+ Hits         1727     1729    +2     
  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.