Emit skainet.tensor_encodings module attribute (#477)

[michalharakal]

Closes #477.

Summary

Promotes the per-op `tensor_encoding` annotation comments introduced in #473 to a structured module-level MLIR attribute that enumerates every encoded tensor in one place. Downstream tools (IREE, any MLIR pass reading SKaiNET output) can now read the encodings via one attribute lookup instead of string-matching against scattered comments.

Emitted output

For a graph with two encoded weight inputs (Q8_0 and Q4_K):

```mlir
module attributes {skainet.tensor_encodings = {w_q4 = "Q4_K", w_q8 = "Q8_0"}} {
func.func @quant_chain(...) -> (...) {
...
}
}
```

Dense graphs with no encoding metadata continue to emit the bare `module {` header byte-for-byte-identically to today — the promotion only fires when the collect phase finds at least one non-null encoding.

Two commits

1. Failing test — `TensorEncodingsModuleAttributeTest`

• `encoded_weights_produce_module_attributes_block` builds a graph with two weight inputs (`TensorEncoding.Q8_0` and `TensorEncoding.Q4_K` via `withTensorEncoding`), runs the converter, and asserts the module header contains `module attributes`, `skainet.tensor_encodings`, and both `w_q8 = "Q8_0"` / `w_q4 = "Q4_K"` entries. Red against pre-fix emitter.
• `dense_graph_keeps_bare_module_header` asserts a dense graph preserves `module {` exactly and introduces no spurious attributes block. A `null` `tensorEncoding` is the unknown / not-carried state, intentionally distinct from `TensorEncoding.Dense`. Green baseline.

2. The fix

• `StableHloConverter.collectTensorEncodings` walks the topologically-ordered node list once, pulling every non-null `tensorEncoding` off each node's input and output specs into a sorted `name -> TensorEncoding` map. First-writer-wins on duplicates.
• `StableHloConverter.convert` emits `module attributes {skainet.tensor_encodings = {...}} {` when the map is non-empty, and `module {` otherwise.
• `MlirValidator` learned three things: (a) module-header lines are not SSA assignments even when they contain ` = ` inside a dict (syntax pass skips them), (b) same for the semantic pass's defined-values extraction, (c) `validateModule` accepts both `module {` and `module attributes` as valid module declaration forms.

Why a module attribute, not per-op

Real per-op MLIR attributes (`stablehlo.dot_general %a, %b {skainet.encoding = "Q8_0"} : ...`) would need every converter in the registry updated — 7 files today, all hand-building their op strings. A module-level attribute lives in exactly one hook and is the biggest value-per-line-of-code first move. Per-op attributes are a better follow-up once the registry is reworked to emit via a builder instead of string concatenation.

Test plan

• `TensorEncodingsModuleAttributeTest` — both cases green
• Full `:skainet-compile:skainet-compile-hlo:jvmTest` — green, no regressions from either the emitter or validator edits
• CI: full multiplatform build

Out of scope

• Per-op MLIR attributes (`stablehlo.*` op custom-assembly-format attributes).
• Real `quant.` dialect emission (`!quant.uniform<i8:f32, 0.1:128>`). Requires structured scale / zero-point parameters that `TensorEncoding` doesn't yet carry.
• Teaching IREE to consume the attribute. Downstream work.
• Removing the per-op `tensor_encoding` comments from Emit TensorEncoding into StableHLO output (P0-1 step 2) #473. They cost nothing now and help human diff readers; can be deleted in a follow-up if we decide the attribute alone suffices.

🤖 Generated with Claude Code