Fix reltol type instability in Newton solver

[ChrisRackauckas-Claude]

Summary

• Fixes type instability in the Newton solver when handling reltol parameter
• Addresses performance issue identified in Type-instability in Jacobian calculation #2819

Problem

The Newton solver had a type instability where reltol could be either:

• A scalar (eps(eltype(dz))) when adaptive=false
• A scalar or vector (integrator.opts.reltol) when adaptive=true

Since adaptive is checked at runtime, Julia couldn't determine the type at compile time, causing performance degradation visible in profiling.

Solution

Changed the condition from:

if integrator.opts.adaptive
    reltol = integrator.opts.reltol
else
    reltol = eps(eltype(dz))
end

To:

if \!integrator.opts.adaptive && \!(integrator.opts.reltol isa AbstractArray)
    reltol = eps(eltype(dz))
else
    reltol = integrator.opts.reltol
end

This ensures reltol is only set to eps(eltype(dz)) when it would be type-compatible (i.e., when the original reltol is not an array), avoiding the type instability.

Test plan

• Tested with the Robertson equation example from issue Type-instability in Jacobian calculation #2819
• Verified the solution works with both adaptive=true and adaptive=false
• Code runs without errors and produces correct results

Fixes part of #2819

🤖 Generated with Claude Code