Change linear op rewrite pattern

[ddilbazTT]

Ticket

#5813
#6016

Problem description

During uplift 2 weeks ago, default tt-metal/ttnn linear op behaviour has changed. linear op is now mapping PyTorch Linear This means that linear op bias needs to be 1d (or view, like 1x1x64)

@ctodTT relaxed the conditions to linear op workaround so that all linear ops would be broken down into matmul and add. However, this means no linear op is running as a linear op, even if they have 1d bias.

What's changed

Linear op workaround is changed such that:

• only linear ops with non 1d bias or batched b input are broken down into matmul + add
• activation is passed from linear to matmul op correctly (previously, this was not passed)
  Tests marked unsupported or skipped are reinstated.

Checklist

[codecov-commenter]

Codecov Report

❌ Patch coverage is 73.80952% with 11 lines in your changes missing coverage. Please review.
✅ Project coverage is 69.38%. Comparing base (0e3246f) to head (05c7744).
⚠️ Report is 24 commits behind head on main.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
...rkarounds/Decomposition/LinearOpRewritePattern.cpp	73.80%	11 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #6014      +/-   ##
==========================================
+ Coverage   69.32%   69.38%   +0.06%     
==========================================
  Files         330      333       +3     
  Lines       50290    50951     +661     
==========================================
+ Hits        34862    35353     +491     
- Misses      15428    15598     +170     

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

[ddilbazTT]

@sdjordjevicTT @jserbedzijaTT I believe this PR would address missing activations in linears with activations. However, I believe you might also consider changing the linear with activation fusing pattern to absorb the repeat that comes before. That way, they would remain as linear with 1d bias.