Add concat / slice / cast StableHLO converters (#489)

[michalharakal]

Closes #489.

Summary

Adds the last generic structural / type tensor ops missing from mainline SKaiNET's StableHLO emitter: concat (plus `concatenate` / `cat` / `stack` aliases), slice, and cast (plus `convert` / `to` aliases). None of these are LLM- or transformer-specific; they're the standard building blocks that sit alongside the existing reshape / flatten / squeeze / unsqueeze coverage.

Emitted lowerings

```mlir
// concat
%out = stablehlo.concatenate %a, %b, dim = :

// slice
%out = stablehlo.slice %x
{start_indices = [..], limit_indices = [..], strides = [..]}
:

// cast
%out = stablehlo.convert %x : () ->
```

• `concat` reads `axis` or `dim` (default 0), normalizes negative axes against rank.
• `slice` reads `start_indices` / `limit_indices` / `strides` (also accepts `starts` / `limits`) and defaults strides to 1 per dim when absent. Static slice only — dynamic / runtime bounds are out of scope.
• `cast` relies on the output spec's dtype and emits the standard MLIR type-transition signature.

Two commits

1. Failing test

`ConcatSliceCastConverterTest` with 7 cases covering all three ops plus every alias.

2. The fix

Extends `ShapeOperationsConverter` (for concat / slice — structural ops belong with the reshape family) and `MathOperationsConverter` (for cast — elementwise primitive). Also updates `ShapeOperationsConverterTest.testSupportedOperations` and `testRegistryIntegration` to cover the new ops — the prior assertions pinned exact set equality on the four reshape ops and would otherwise regress on this additive change. That regression was caught locally by `allTests` across all targets, not by my earlier narrowly-filtered `jvmTest --tests=ConcatSliceCastConverterTest` run.

Test plan

• `ConcatSliceCastConverterTest` — 7/7 green
• `ShapeOperationsConverterTest` — updated, green
• `./gradlew :skainet-compile:skainet-compile-hlo:allTests -x kotlinWasmStoreYarnLock` — green across jvmTest, wasmJsTest, wasmJsBrowserTest, wasmWasiTest, wasmWasiNodeTest, macosArm64Test, iosSimulatorArm64Test (`linuxX64Test` skipped on macOS host)
• CI: full multiplatform build

Out of scope

• Dynamic slice with runtime bounds. Static slice is sufficient for the current emitter targets; can be added when a traced model surfaces the pattern.
• `gather` / `scatter` — separate converter, gather already landed in Add gather / embedding StableHLO converter (#483) #487.
• Quantization-aware cast that preserves `TensorEncoding`. The current converter just maps element types; a quant-preserving variant can come in a follow-up.

Roadmap context

This closes out the generic-op coverage needed before the IREE compile-path spike. With this merged, the mainline StableHLO emitter can reasonably handle a full transformer forward pass up to the engine boundary — any LLM-specific ops beyond this point belong in `SKaiNET-transformers`.

🤖 Generated with Claude Code