Decompose TurboQuant into 2 ScalarFnArrays and DictArray

[connortsui20]

Summary

Tracking issue: #7297

Decomposes TurboQuant into:

ScalarFnArray(L2Denorm, [
    ScalarFnArray(SorfTransform, [
        Extension<Vector>(
            FixedSizeListArray(
                DictArray(codes=Primitive<u8>, values=Primitive<f32>),
                padded_dim
            )
        )
    ]),
    norms
])

This makes the implementation more modular and turns the TurboQuant-specific pieces into reusable building blocks.

Also defines SORF sign generation with a frozen local SplitMix64 implementation and cleans up related vector compute code.

API Changes

• add SorfTransform
• remove the TurboQuant encoding/array type
• keep TurboQuantScheme as the compressor entry point
• simplify a all scalar fn APIs as we no longer need ApproxOptions

Note that ApproxOptions doesn't actually make sense here because we are doing exact compute here, it is the encoding itself that is lossy. Until we figure out what the exact semantics are of a lossy encoding I will remove this.

Testing

More tests:

• TurboQuant roundtrip / structural coverage
• SORF roundtrip + norm preservation
• deterministic SplitMix64 / sign generation coverage
• readthrough behavior (when children are normalized) for vector similarity ops

[codspeed-hq]

Merging this PR will not alter performance

✅ 1122 untouched benchmarks
⏩ 1455 skipped benchmarks¹

---

_{Comparing ct/tq-decomp (c0f6037) with develop (360b41d)}

Footnotes

1. 1455 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[gatesn]

This seems wrong to me...

As in, I don't think ScalarFn::execute should be matching on child array type should it? CC @joseph-isaacs

[connortsui20]

we resolved this, this is fine