Fix PSO benchmark : skip unstable BBOB functions

[AdityaPandeyCN]

Pin CUDA runtime via LocalPreferences extras and exclude f4, f7, f10 which cause segfaults or illegal memory access on GPU.

Checklist

• Appropriate tests were added
• Any code changes were done in a way that does not break public API
• All documentation related to code changes were updated
• The new code follows the
  contributor guidelines, in particular the SciML Style Guide and
  COLPRAC.
• Any new documentation only uses public API

Additional context

The PR also adds updated Manifest.toml

[ChrisRackauckas]

What's the cudaruntime thing? That's probably not the way to solve it.

Also, why was the benchmark okay with these before but not now?

[AdityaPandeyCN]

What's the cudaruntime thing? That's probably not the way to solve it.

I aimed to resolve this failure from earlier benchmark run

ERROR: Compute capability 7.0.0 is not supported by CUDA 13.2.0
Stacktrace:
  [1] (::Weave.var"#32#34"{String, Module, String, Bool, Task, Base.PipeEndpoint, Base.PipeEndpoint, Base.PipeEndpoint})()
    @ Weave ~/.julia/packages/Weave/f7Ly3/src/run.jl:222
  [2] task_local_storage(body::Weave.var"#32#34"{String, Module, String, Bool, Task, Base.PipeEndpoint, Base.PipeEndpoint, Base.PipeEndpoint}, key::Symbol, val::String)
    @ Base ./task.jl:304
  [3] capture_output(code::String, mod::Module, path::String, error::Bool, report::Weave.Report)
    @ Weave ~/.julia/packages/Weave/f7Ly3/src/run.jl:216
  [4] #29
    @ ~/.julia/packages/Weave/f7Ly3/src/run.jl:191 [inlined]
  [5] _broadcast_getindex_evalf
    @ ./broadcast.jl:709 [inlined]
  [6] _broadcast_getindex
    @ ./broadcast.jl:682 [inlined]
  [7] getindex
    @ ./broadcast.jl:636 [inlined]
  [8] macro expansion
    @ ./broadcast.jl:1004 [inlined]
  [9] macro expansion
    @ ./simdloop.jl:77 [inlined]
 [10] copyto!
    @ ./broadcast.jl:1003 [inlined]
 [11] copyto!
    @ ./broadcast.jl:956 [inlined]
 [12] copy
    @ ./broadcast.jl:928 [inlined]
 [13] materialize(bc::Base.Broadcast.Broadcasted{Base.Broadcast.DefaultArrayStyle{1}, Nothing, Weave.var"#29#30"{Weave.Report, Module, Bool, String}, Tuple{Vector{String}}})
    @ Base.Broadcast ./broadcast.jl:903
 [14] run_code(doc::Weave.WeaveDoc, chunk::Weave.CodeChunk, report::Weave.Report, mod::Module)
    @ Weave ~/.julia/packages/Weave/f7Ly3/src/run.jl:192
 [15] eval_chunk(doc::Weave.WeaveDoc, chunk::Weave.CodeChunk, report::Weave.Report, mod::Module)
    @ Weave ~/.julia/packages/Weave/f7Ly3/src/run.jl:263
 [16] run_chunk(chunk::Weave.CodeChunk, doc::Weave.WeaveDoc, report::Weave.Report, mod::Module)
    @ Weave ~/.julia/packages/Weave/f7Ly3/src/run.jl:133
 [17] run_doc(doc::Weave.WeaveDoc; doctype::String, out_path::String, args::Dict{Symbol, String}, mod::Nothing, fig_path::Nothing, fig_ext::Nothing, cache_path::String, cache::Symbol)
    @ Weave ~/.julia/packages/Weave/f7Ly3/src/run.jl:74
 [18] weave(source::String; doctype::String, informat::Nothing, out_path::String, args::Dict{Symbol, String}, mod::Nothing, fig_path::Nothing, fig_ext::Nothing, cache_path::String, cache::Symbol, template::Nothing, css::Nothing, highlight_theme::Nothing, pandoc_options::Vector{String}, latex_cmd::Vector{String}, keep_unicode::Bool)
    @ Weave ~/.julia/packages/Weave/f7Ly3/src/Weave.jl:176
 [19] weave
    @ ~/.julia/packages/Weave/f7Ly3/src/Weave.jl:124 [inlined]
 [20] weave_file(folder::String, file::String, build_list::Tuple{Symbol, Symbol})
    @ SciMLBenchmarks ~/_work/SciMLBenchmarks.jl/SciMLBenchmarks.jl/src/SciMLBenchmarks.jl:61
 [21] top-level scope
    @ none:5

caused by: LoadError: Compute capability 7.0.0 is not supported by CUDA 13.2.0

Also, why was the benchmark okay with these before but not now?

It ran for me on T4 colab GPU but for utkarsh's cluster run of this benchmark these functions were problem, precisely this.
Affected (3/19) F4 Buche-Rastrigin, F10 Ellipsoidal-2 (catchable illegal-access), F7 Step-Ellipsoidal (hard segfault)

[AdityaPandeyCN]

@ChrisRackauckas Benchmarks are finally reproduced by the CI :)

[AdityaPandeyCN]

@ChrisRackauckas Can you take a look at the artifacts?
https://github.com/SciML/SciMLBenchmarks.jl/actions/runs/29597613342/artifacts/8435117115

I probably need more tuning but let me know your views on the graphs like how you want them to be?

[ChrisRackauckas]

Why are all of the success rates so low now?

BBO was basically 100%, now it's not even 50%. That's a massive regression, it obviously cannot be merged as-is.

[AdityaPandeyCN]

I increased the dimension of the BBO test functions to 10 from 3, which has increased the difficulty and caused the regression.
I will do some tuning and runs to get the best of pso before we can merge

[ChrisRackauckas]

No please follow the procedure you would do with any other PR and isolate the changes. Do one PR that is just adding the new methods. Then one PR that is increasing the difficulty. Step by step changes where provenance is clear.

[AdityaPandeyCN]

Apologies, have reset to the relevant commit.