Convert nlsolver.γ to tTypeNoUnits to avoid expensive rationalize calls

[ChrisRackauckas-Claude]

Summary

This PR fixes #2921 by converting the nlsolver.γ parameter to tTypeNoUnits in build_nlsolver, preventing expensive rationalize() calls during QNDF integration.

Problem

When using QNDF solvers, the nlsolver.γ was initialized as Rational{Int64} (e.g., Int64(1)//1). Later, during the step function, a Float64 value would be assigned to nlsolver.γ = β₀. This triggered Julia's conversion from Float64 to Rational{Int64}, which calls rationalize() - an expensive operation that was consuming significant computation time.

Solution

Convert γ to tTypeNoUnits (typically Float64) in build_nlsolver before:

1. Using it in nlp_params tuples for NonlinearSolveAlg
2. Constructing the NLSolver object

This ensures the type parameter gamType is Float64 from the start, so later assignments are just Float64 to Float64 (no conversion needed).

Changes

• lib/OrdinaryDiffEqNonlinearSolve/src/utils.jl: Added γ = tTypeNoUnits(γ) conversion in both Val{true} (in-place) and Val{false} (out-of-place) versions of build_nlsolver

Benchmark Results

using OrdinaryDiffEqBDF, BenchmarkTools
f(du, u, p, t) = du .= 0.0
prob = ODEProblem(f, ones(10), (0.0, 1.0))
@btime solve(prob, QNDF(); adaptive = false, dt = 1e-5, save_everystep = false)

• Before: ~391 ms
• After: ~264 ms (~32% speedup)

Test Plan

• Pkg.test("OrdinaryDiffEqBDF") passes (70 tests)
• Pkg.test("OrdinaryDiffEqNonlinearSolve") passes
• Pkg.test("OrdinaryDiffEqSDIRK") passes
• Manual verification that QNDF, QNDF1, QNDF2, and FBDF solvers work correctly

Closes #2921

🤖 Generated with Claude Code