Fix st_cast empty geometry split ordering

[kendonB]

Summary

• preserve empty geometries when st_cast splits multiparts into multiple rows
• rebuild output in original order and keep ids so attributes repeat correctly

Testing

• devtools::load_all('.')
• ran user repro (MULTIPOLYGON with empty) -> returns 5 rows with correct names/bboxes

[kendonB]

FYI this was all gpt-5.1-codex-max-xhigh (including the PR)

[edzer]

That's a lot of code. Could you provide a very small example of what this code tries to solve?

[rsbivand]

Testing should also include full reverse dependency checks of all reverse depencencies, with comparative timings and memory profiling.

[kendonB]

Sorry I missed the reference to the issue which contains the small example of what this code tries to solve #2555

[kendonB]

Testing should also include full reverse dependency checks of all reverse depencencies, with comparative timings and memory profiling.

Running this now with r-lib/revdepcheck

[edzer]

Did the revdepcheck finish?

[kendonB]

Did the revdepcheck finish?

It did but it didn't manage to check everything.

## Failed to check (10)

|package     |version |error |warning |note |
|:-----------|:-------|:-----|:-------|:----|
|arcgisutils |0.4.0   |1     |        |     |
|awdb        |0.1.3   |1     |        |     |
|excursions  |2.5.11  |1     |        |     |
|GeoFIS      |1.1.1   |1     |        |     |
|h3o         |0.3.0   |1     |        |     |
|hgwrr       |0.6-2   |1     |        |     |
|landsepi    |1.5.2   |1     |        |     |
|linkeR      |0.1.3   |1     |        |     |
|protolite   |2.3.1   |1     |        |     |
|socratadata |0.1.1   |1     |        |     |

## New problems (3)

|package |version |error  |warning |note |
|:-------|:-------|:------|:-------|:----|
|[btb](problems.md#btb)|0.2.1   |__+1__ |        |     |
|[r5r](problems.md#r5r)|2.3.0   |__+1__ |        |     |
|[SHARK4R](problems.md#shark4r)|1.0.2   |__+1__ |        |     |

The new problems are also new problems on main (vs CRAN). Do you consider this a pass?

I haven't done memory profiling or comparative timings yet, but wouldn't a more targeted performance test (testing only cast) be better?