chore(linter): add TRF018 modeling rule (#46259)
* Add TRF018 modeling rule
* Fix TRF018 init weights batch 1
* Fix TRF018 init weights batch 2
* Fix TRF018 init weights batch 3
* fix style
* ignore rule 18 on certain classes
* fix sam3
* more cleanups
* revert mlinter change
* new pass
* revert rule change
* remove change
* simplify some branches with super
* p1
* part2
* remove ignores where not needed
* i have no idea why but let's keep it ig
* fix
* ignore version-based kernel issue
* make fmt
* Fix regnet
The parent PreTrainedModel._init_weights draws random numbers for every module
before regnet's own Conv2d/Linear override runs — and for Conv/Linear that
parent result is immediately overwritten, so the only effect is shifting the
global RNG stream. The test checks that the original classifier's 3-element
bias (init'd with uniform_, which ignores config_no_init) has mean <= 0.1 — a
high-variance, RNG-dependent check. The shift pushed regnet's draw to mean
0.129 → fail. It's not a behavioral regression, just RNG luck flipping.
---------
Co-authored-by: vasqu <antonprogamer@gmail.com>
