Fold the receptive field into the Xavier fans

[jessegrabowski]

sqrt(2 / sum(shape)) is the right fan computation only for a matrix, so a conv kernel came out about 2.7x too wide, and a 1-D parameter got a number with no meaning behind it — that now raises. Nothing in the tree has a 4-D weight yet, so this lands ahead of the conv work rather than fixing anything shipping, and every existing 2-D parameter draws bit-identically since the fan sum is the shape sum there. Fan-in is the leading dimension, matching Linear's (n_in, n_out) rather than torch's transpose of it — invisible to Xavier, load-bearing for anything that scales by fan-in alone.

Two unrelated riders: Dropout builds its generators with pytensor's shared_rng, and the tracked mnist notebook's matplotlib and tqdm are declared.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.43%. Comparing base (f60deb7) to head (acdc8cc).

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #77      +/-   ##
==========================================
+ Coverage   95.42%   95.43%   +0.01%     
==========================================
  Files          45       45              
  Lines        1792     1798       +6     
==========================================
+ Hits         1710     1716       +6     
  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.