feat(data): add Iris dataset provider (#1044)

[AjithGoveas]

What

Adds a first-class Iris dataset provider to skainet-data-simple, enabling one-line access to the 1936 Fisher Iris dataset for classification tutorials.

val (train, test) = Iris.load().split(0.8, seed = 42L, stratified = true)

Files changed

• Iris.kt — entry object: featureNames, classNames, suspend fun load()
• IrisData.kt — IrisSample data class, embedded 150-row CSV (UCI canonical ordering, public domain), strict parser with error messages naming line + column
• IrisDataset.kt — Dataset<FloatArray, Int>: FP32 [n,4] features, FP32 [n,3] one-hot labels. Both createDataBatch and createIndexedDataBatch overridden so split()/shuffle() views batch correctly
• IrisDatasetTest.kt — 8 tests in jvmTest (same convention as MNIST/CIFAR tests)
• README.md — Iris listed in built-in loaders
• data-sources-getting-started.adoc — Iris usage snippet added

Design notes

• Data is source-embedded (no file resource) to stay platform-agnostic including JS and Wasm browser targets, matching the repo's existing approach
• load() is suspend for call-site symmetry with MNIST.loadTrain() / CIFAR10.load()
• Raw centimetres returned; no standardisation (scaling should be fitted on training split only)
• getY() returns Int, not one-hot — one-hot conversion is deferred to batch construction so stratified splitting works correctly

Checklist

• Closes [Feature]: Add an Iris dataset provider (and a tabular RawDataset -> Dataset bridge) #1044
• All 8 Iris tests pass
• No regressions in existing skainet-data-simple tests
• Passes explicit-api compilation
• Public API has KDoc

[michalharakal]

First of all: welcome, and thank you, the quality bar you set here is genuinely impressive.
The createIndexedDataBatch override for non-contiguous views, deferring one-hot to batch construction so stratified splitting works, and the test proving the split-then-batch path — that's exactly the subtle stuff that's usually missed. 🎉

I pushed one tiny fix directly to your branch (0dd4620) so we can merge without another round-trip: in the unknown-species error message, $Iris.classNames interpolated Iris.toString() plus the literal text .classNames — it needed braces, ${Iris.classNames}. I also added an assertion to the parser test so it would catch this in future.

Two purely optional, non-blocking notes — fine as follow-ups or to just leave as-is:

1. The no-arg shuffle() override uses Random.Default, so it's non-deterministic. Probably intentional, but a short KDoc note pointing reproducibility-minded users at the seeded shuffle(seed) overload would help.
2. The plain split(splitRatio) cuts at a raw index, and since the embedded data is ordered by species (50 per class), an unstratified 0.8 split puts every virginica into the test set. A KDoc warning recommending split(ratio, seed, stratified = true) would save someone a confusing afternoon.

Thanks again — looking forward to more! I am preparing another similar first good issues so we can reach funstion/Ops parity with KotlinDL :-) #290