Fix softmax StableHLO lowering to use real reductions

[michalharakal]

Problem

ActivationOperationsConverter.convertSoftmax() currently emits numerically incorrect StableHLO. It hardcodes placeholder constants instead of actual reductions:

%maxValue  = stablehlo.constant dense<0.0> : tensor<?xf32>   // should be reduce(max)
%shifted   = stablehlo.subtract %input, %maxValue
%exp       = stablehlo.exponential %shifted
%sumValue  = stablehlo.constant dense<1.0> : tensor<?xf32>   // should be reduce(add)
%result    = stablehlo.divide %exp, %sumValue

The max(x) and sum(exp(...)) terms are replaced with constants, so:

• Every softmax output is mathematically wrong.
• Any model exported through StableHloConverter with a softmax is unusable downstream (IREE, MLIR tools, NPU compile path).
• This is especially blocking for the LLM → StableHLO → IREE → NPU (int8/int4) path, where softmax is on the critical path inside every attention block.

File: skainet-compile/skainet-compile-hlo/src/commonMain/kotlin/sk/ainet/compile/hlo/converters/ActivationOperationsConverter.kt (around lines 92-150).

Expected lowering

// shape-preserving softmax along `axis`
%neg_inf = stablehlo.constant dense<-3.4028235e+38> : tensor<f32>
%max     = stablehlo.reduce(%input init: %neg_inf) applies stablehlo.maximum across dimensions = [axis]
%max_b   = stablehlo.broadcast_in_dim %max, dims = [non-axis dims] : tensor<...>
%shifted = stablehlo.subtract %input, %max_b
%exp     = stablehlo.exponential %shifted
%zero    = stablehlo.constant dense<0.0> : tensor<f32>
%sum     = stablehlo.reduce(%exp init: %zero) applies stablehlo.add across dimensions = [axis]
%sum_b   = stablehlo.broadcast_in_dim %sum, dims = [non-axis dims] : tensor<...>
%result  = stablehlo.divide %exp, %sum_b

Scope

• Rewrite convertSoftmax to emit the real reduction form above.
• Handle the axis parameter correctly (negative axes normalized against rank).
• Unit test in ActivationOperationsConverterTest: assert no dense<0.0> / dense<1.0> placeholder constants are emitted, and that two stablehlo.reduce blocks are present.
• Update any golden MLIR in StableHloExportTest or round-trip tests that was accepting the wrong form.

Out of scope

• RMSNorm converter (separate issue).
• LayerNorm proper lowering (currently custom_call, separate issue).
• Quantized softmax / int8 attention (depends on P0 quant-in-IR work).

Context

Part of the priority-ordered NPU/IREE roadmap. This is P1 (core compile path, correctness prerequisite) — cheap, high-signal, unblocks every transformer export.