Add Flash IPA

[AntonOresten]

Adds an experimental version of Flash IPA that doesn't use factorized pair representations, but instead only uses the full $O(L^2)$ pair features to create the pair bias, that gets passed into an attention function that supports it.

The "head_dim" is not the same dim that gets used in the MHA. That instead becomes head_dim + (3+1+1)*num_query_points, so if the attention function is a flash attention that only supports dimensions that are power of two, the hidden size may need to be odd-ish. Might also need to distinguish between $d_{qk}$ and $d_v$.

In pseudo-pseudo-code, it should roughly be:

[codecov]

Codecov Report

❌ Patch coverage is 0% with 35 lines in your changes missing coverage. Please review.
✅ Project coverage is 29.60%. Comparing base (79ab829) to head (9ea6bc1).

Files with missing lines	Patch %	Lines
src/ipa/FlashIPA.jl	0.00%	35 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #47      +/-   ##
==========================================
- Coverage   30.79%   29.60%   -1.20%     
==========================================
  Files          43       44       +1     
  Lines         867      902      +35     
==========================================
  Hits          267      267              
- Misses        600      635      +35     

☔ View full report in Codecov by Sentry.
📢 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.