Expose all kwargs to optimize_acqf in optimize_acqf_homotopy

[CompRhys]

Motivation

This PR seeks to address #2579

Have you read the Contributing Guidelines on pull requests?

Yes

Test Plan

Added test incorporating a linear inequality constraint

Related PRs

If this is merged as is with breaking changes then a follow-on PR will be needed in Ax

[saitcakmak]

As discussed on the issue, this is quite a big divergence from the signature of the other optimize_acqf implementations. Rather than diverging, we want to unify the signatures as much as possible, so that the code that uses one optimizer could easily switch over to another.

In addition, dict valued inputs are generally not desirable. They do not contain any meaningful type hints, which provides very limited information to the user to what the inputs should be and requires them to dig through the codebase to figure out what is acceptable. We want to reserve these only for rarely modified options.

[codecov]

Codecov Report

Attention: Patch coverage is 67.85714% with 9 lines in your changes missing coverage. Please review.

Project coverage is 99.93%. Comparing base (d2c6f5e) to head (85805d9).
Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
botorch/optim/optimize_homotopy.py	67.85%	9 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2580      +/-   ##
==========================================
- Coverage   99.98%   99.93%   -0.06%     
==========================================
  Files         196      196              
  Lines       17333    17357      +24     
==========================================
+ Hits        17331    17346      +15     
- Misses          2       11       +9     

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

[CompRhys]

I really am loath to contribute a PR that creates increased maintenance burden going forward for sake of documentation/type hinting, this method has these missing constraints options and has the apparent fixed_features bug because it leaves the surface area for such issues to exist by hard coding the args to the function it wraps. There will 100% be divergences again when more functionality is added to the base method.

However, pragmatically as I would like this functionality, if you get back to me with decisions about the need for final_options (I think it's most consistent with what you propose in terms of arg unification to remove this as not used here or in Ax) and whether the apparent issues with fixed_features are indeed bugs or whether it was intended (I can't see how it would be intended as the test would be flaky if the test optimization was harder), I can implement it the other way.

test failures are all saying that 2.000 is not >= 2.0 which I can debug after re-writing the wrapper function

[CompRhys]

close in favour of #2588