[MINOR] `MLEStrategy` lacks friendly validation of empty / non-finite sample arrays (cryptic internal errors)

[MINOR] MLEStrategy lacks friendly validation of empty / non-finite sample arrays (cryptic internal errors)

Summary

MLEStrategy (src/strategies/mle.jl) performs no input validation on the sample array. Degenerate inputs surface as cryptic internal errors instead of a clear user-facing message, after the solver has already started.

Evidence (Julia 1.12.6)

• project_to(ProjectedTo(Beta), Float64[]) throws ArgumentError: reducing over an empty collection is not allowed; consider supplying init (raised deep inside Statistics.mean in MLETargetFn at src/strategies/mle.jl:89).
• project_to(ProjectedTo(NormalMeanVariance), [NaN]) throws LinearAlgebra.PosDefException (from cholinv(fisherinformation(...)) on a NaN-poisoned iterate).

Neither error message tells the caller their input array was empty or contained NaN/Inf.

Impact

Minor — these are degenerate inputs — but the poor error surfacing can be genuinely confusing (e.g. in ReactiveMP/RxInfer wiring where an empty message array can occur through a bug elsewhere).

Suggested fix

Validate in project_to (or MLEStrategy.create_state!): reject empty and non-finite samples with a clear ArgumentError before optimization starts.

Start in src/strategies/mle.jl, reading MLEStrategy.create_state! and the MLETargetFn code around line 89, then trace the project_to entry point. Confirm that empty or non-finite sample arrays are rejected with clear ArgumentError messages before optimization begins, rather than producing internal solver errors.