feat: first-class dynamic tensor dimensions for streaming KV-cache decode (squashed)

[michalharakal]

Note: This is the single-commit variant of the dynamic-shape capability, cut from the sibling checkout used for -transformers composite builds (on top of the argMax fix, with the SoftmaxCpuTest fix folded in). It is functionally identical to #891 (which has the same change as 5 reviewable commits). Only one of the two should be merged — pick whichever base/history you prefer and close the other.

Summary

Makes dynamic tensor dimensions a first-class, correctly-compiling concept end-to-end, so a single compiled vmfb serves every autoregressive decode position — unbounded streaming KV-cache decode from the pure SKaiNET DSL.

Previously "dynamic decode" existed only as a downstream sentinel-prime + post-emit text hack (x7919x → x?x) that did not actually compile: the emitter produced static op forms that iree-compile rejects once a dim is ?. This PR fixes that at the source and removes the overloading of -1 (which meant "dynamic" in the emitter, "infer" in reshape, and "last index" in the slice DSL).

Closes #890.

What changed

First-class Dim (skainet-lang-core)

• Dim.DYNAMIC = Int.MIN_VALUE — a reserved sentinel distinct from reshape's -1 = infer — plus centralized dynamic arithmetic (concat, compatible, render, isDynamic/isStatic).
• Shape: hasDynamic(), isDynamic(axis), dynamicAxes; volume now throws on a dynamic shape instead of returning a corrupt product.
• Slice DSL: all() over a dynamic axis is a valid symbolic full-axis.

Shape-only, dynamic-safe tracing (skainet-lang-core, skainet-compile-dag)

• VoidTensorOps propagates shapes through a ShapeOnlyTensorData that allocates no backing buffer, so a -1 extent threads through a whole decode trace (was NegativeArraySizeException). Concat routes through Dim.concat; reshape passes a dynamic target through.
• TraceToGraphBuilder.extractFloatArray no longer probes volume of a non-dense/dynamic input.

Dynamic-shape-safe StableHLO emission (skainet-compile-hlo)

• Scale folded into Q ((q·s)@kᵀ ≡ scores·s) — removes the invalid scores-sized dynamic splat constant.
• Softmax reduce/broadcast uses stablehlo.dynamic_broadcast_in_dim + a runtime get_dimension_size shape operand (CHLO is rejected by IREE's stablehlo pipeline).
• Identity reshape, full-extent slice/narrow elided; concat keeps the axis dynamic (? ++ 1 stays ?, never a bogus 0).
• TypeMapper.DYNAMIC_DIM aliases Dim.DYNAMIC so tracer and emitter agree by construction.

Every dynamic path is gated behind hasDynamic(); static graphs are emitted byte-for-byte unchanged.

Verification

• Full skainet-lang-core + skainet-compile-hlo + skainet-compile-dag suites pass; new tests cover Dim arithmetic, guarded volume, dynamic concat/reshape/slice, and dynamic SDPA emission.
• Synthetic dynamic SDPA iree-compiles and runs at key length 3 and 17 from one module.
• FunctionGemma with_past (dynamic 1x{nKV}x?x256, real weights) self-compiles from the DSL → 151 KB vmfb (previously uncompilable).
• Moonshine v2 with_past (dynamic self and cross caches 1x8x?x40, weights baked) self-compiles → 86 MB vmfb.
• Growing-cache greedy decode of the Moonshine v2 dynamic vmfb vs onnxruntime decoder_kv.onnx, cache 1→12: cos = 1.000000 and argmax match at every step.

[github-actions]

📖 Documentation Preview

The documentation has been built successfully for this PR.

Generated Files:

• Operator documentation: docs/modules/operators/_generated_/
• JSON schema output: operators.json

Artifacts:

• Download the documentation-preview-892 artifact to view the complete documentation locally.

This comment will be updated automatically when the PR is updated.