ExponentialRK: @views residual column-slice updates to remove per-step allocations

[ChrisRackauckas-Claude]

Summary

Follow-up to #3928. The EPIRK/Exp4/Exprb steppers still allocated per step
through bare column-slice compound assignments of the form X[:, i] .op= ...
(e.g. K[:, i] ./= ts[i], B[:, 4] .-= c .* rtmp). Without @views, the
right-hand-side read of the slice materializes a fresh length-n column copy
each time (Base _unsafe_getindex -> similar), so each such line allocates one
working vector per step and the cost grows linearly with the state size.

#3928 converted the large fused updates to @views @.. broadcast=false but
left these smaller normalization/update slices unviewed. This wraps the
remaining 17 of them in @views. Each is an in-place elementwise write back to
the same slice with a scalar or separate-array RHS (no column aliasing), so the
result is unchanged — only the transient copy is removed.

Measured effect

Per-step allocations via the step! interface after warmup, on a 1D semilinear
reaction–diffusion problem (N=128, adaptive_krylov, m=30), master vs this
branch:

method	master	this PR	reduction
Exp4	7647 B	927 B	8.2×
EPIRK4s3B	5132 B	652 B	7.9×
EPIRK4s3A	2879 B	639 B	4.5×
EXPRB53s3	3180 B	940 B	3.4×
EPIRK5P1	2047 B	927 B	2.2×

The final solution is bit-identical before/after for every method
(max|Δu| = 0 across Exp4/EPIRK4s3A/EPIRK4s3B/EXPRB53s3/EPIRK5P1/EPIRK5P2/
Exprb32/Exprb43).

Test

Adds a runtime size-independence regression test: Exp4's per-step allocation
must not scale with the state size n (measured at n=32 vs n=256). Exp4
uses the symmetric-Jacobian Lanczos path with a stable Krylov subspace, so once
the slices are views its per-step allocation is n-independent; a reintroduced
slice copy makes it scale with n and trips the test. This is version- and
platform-robust, unlike an absolute byte ceiling.

Verified locally: the Core group (Linear-Nonlinear Krylov + Convergence
tests) passes; the new regression test passes (ratio ≈ 1.0 on this branch, ≈
4.8 on master). The 3 pre-existing Aqua/QA failures (missing EPIRK docstrings

• reexport hygiene) are unrelated to this change and already fail on master.

Remaining work (separate PR)

At large n, the EPIRK family still allocates in the ExponentialUtilities
adaptive phiv_timestep! path — the matrix-exponential Padé workspace is
reallocated (alloc_mem) when the adaptive subspace size varies (cache keyed by
size → misses), and the Krylov subspace is resize!d. That is a separate
ExponentialUtilities change and is not addressed here.

---

Draft — please ignore until reviewed by @ChrisRackauckas.

🤖 Generated with Claude Code