Give the Newton-Krylov round-off test explicit tolerances

[singhharsh1708]

Fixes #4293. Test only, no library change, per @ChrisRackauckas on #4301.

newton_krylov_roundoff.jl called solve with no tolerances. The linear solve inherits
reltol, so the Krylov solve ran at the default 1e-3 and returned a Newton direction accurate
only to 1e-3, which cannot drive the stage to 1e-3. dt collapses and both methods return
Unstable.

The part worth checking is that the test had stopped discriminating. At the default tolerance it
fails identically whether the warm start is the guarded one that #4040 shipped or the
unpredictive reuse that #4034 reported, so it was not testing anything:

warm start	default tolerance	reltol = abstol = 1e-12
guarded Hegedüs, current default	Unstable	Success, drift 6.0e-07 and 4.6e-07
WarmStart.Previous, regression proxy	Unstable	Unstable

At 1e-12 the good configuration passes and the bad one still fails, so the gate protects what it
was written to protect.

On picking 1e-12, measured on this DAE with maxiters = 10_000 left as it was:

tolerance	Hairer4	Hairer42
default 1e-3	Unstable	Unstable
1e-7	drift 2.0e-04	drift 5.1e-05
1e-8	drift 3.1e-05	drift 3.4e-05
1e-9	drift 1.1e-06	drift 2.7e-05
1e-10, 1e-11	MaxIters	MaxIters
1e-12	drift 6.0e-07, 1211 steps	drift 4.6e-07, 1326 steps

1e-7 through 1e-9 complete but sit at or above the 1e-5 gate, so they would not be a working
test. 1e-12 clears the gate by a factor of 16 and uses 1211 and 1326 of the 10,000 step budget.
The non-monotonic MaxIters at 1e-10 and 1e-11 is the chaotic step sequence the file's comment
already describes.

The comment block is updated to say why the tolerances are explicit and to quote the drift
actually measured, since the old text described a 1e-15 to 1e-7 band that no longer matches.

Verified by running the file against unmodified master at 13da423: 3/3 and 3/3, exit 0.

AI Disclosure

Claude assisted with this work.

[singhharsh1708]

Duplicate of #4335, which I missed before opening this. #4335's tolerances are also better: drift 2.9e-09 and 1.5e-09 against the 1e-5 gate, where mine sat at 6.0e-07.