fix(translate): drop identically-zero terms to fix the correlation regression

[oameye]

The benchmark CI was right after all: there's a real regression in the Correlations/two-time (and time-dependent interaction-picture) benchmark, and it's not cross-runner noise. I tracked it down to translate_qo.

The time-dependent path kept terms whose coefficient is identically zero once the parameters are substituted. When a coupling is set to zero, the substitution leaves a conj(0.0) factor (or conj(conj(0.0)) when the cascade Hamiltonian's second cross-term is built as adjoint(X)), and Symbolics doesn't fold either back to 0. So the term looked nonzero, survived into the compiled closure, and got re-evaluated (pulse interpolation + operator rebuild) on every single ODE step. The g_v => 0 coupling in the correlation benchmark hits this directly.

The fix folds conj of a numeric literal in each coefficient and drops the ones that collapse to zero. conj of a symbolic time variable is left alone, so genuinely time-dependent complex couplings are untouched.

Measured on the correlation benchmark (same machine, g_v = 0):

	allocations	memory	per-call H(t)
before (current sqav0.6)	5.17M	555 MB	6128 B
after this PR	2.78M	373 MB	2080 B
SQA 0.4.5 (old baseline)	2.78M	373 MB	2400 B

So it's back to the old 0.4.5 allocation level (actually a hair below per-call). Output is bit-identical (dropping identically-zero terms is exact), numeric SLH composition is unchanged since this doesn't touch the cascade, and translate construction time is unchanged. test_translate (37) and test_correlations (4) pass locally.

Worth noting: the earlier "it's just runner noise" framing on PRs #15/#18 was wrong, this was a genuine ~1.5-2x slowdown that the minimum/gcsample work correctly surfaced instead of hiding.

[oameye]

Superseded by the proper fix in SecondQuantizedAlgebra: qojulia/SecondQuantizedAlgebra.jl#193.

The dead terms this PR dropped in translate_qo came from SQA leaving conj(conj(x)) (cascade adjoint) and conj(<const>) (a coupling substituted to zero) unfolded. SQA #193 folds both at the source, so substitute drops the zero terms before translation sees them. Verified the exact shapes here (time-dependent g_u times a conj-of-zero coupling) collapse to 0, so this QIO-side workaround is no longer needed.

Closing rather than patching QIO. Once SQA #193 is released, QIO just needs a [compat] bump to pick it up. (Reopen if we want a temporary bridge to green CI before that release.)