Use Some/Nothing instead of a sentinel for fzero

[LilithHafner]

Fixes #54467

[jishnub]

Overall, this looks good to me, but the commit message should be a bit more informative.

[KristofferC]

Would be good with a quick spot check to see that performance is not impacted.

[LilithHafner]

but the commit message should be a bit more informative.

Yes, I usually write those upon squash-merging. Perhaps

Use Some/nothing instead of a missing as sentinel for fzero to avoid bug that destroys sparsity when missing occurs in input

[LilithHafner]

@nanosoldier runbenchmarks("broadcast")

[nanosoldier]

Your benchmark job has completed - successfully executed benchmarks. A full report can be found here.

[LilithHafner]

@nanosoldier runbenchmarks("broadcast", vs=:master)

[LilithHafner]

@nanosoldier runbenchmarks("broadcast", vs=":master")

[KristofferC]

I don't know if the broadcast category in BaseBenchmarks contains anything using this code.

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here.

[LilithHafner]

Do you have any reason to think this could impact performance?

@nanosoldier runbenchmarks(ALL, vs=":master")

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here.

[LilithHafner]

Don't see any regressions from nanosoldier.

[LilithHafner]

@b Diagonal(rand(3)) _ .+ _ also exercises this code path and does not regress.

[KristofferC]

Do you have any reason to think this could impact performance?

The broadcast code is very inference sensitive and I've seen many small changes cause regressions due to e.g. something suddenly stop to inline.