Add TurboQuant Swift support primitives

[RNT56]

Summary

Adds the lower-level Swift TurboQuant support used by the LM-side KV-cache work:

• packed TurboQuant tensor representation and reference codec
• PolarQuant/QJL Metal codec and compressed-attention support paths
• SwiftPM-built default.metallib resource for package test/runtime contexts
• runtime capability probing and fallback behavior
• focused TurboQuant tests, including runtime probe and packaged-metallib checks

This is opened as a draft because it is the support branch for the LM-side split and may need to move depending on the preferred boundary with MLX/custom kernels.

Validation

• git submodule update --init --recursive
• swift test --filter TurboQuant
  • passed: 23 selected tests, 0 failures
  • verifies the SwiftPM Metal library resource is built/copied and runtime TurboQuant Metal paths execute
• swift test
  • passed: 543 tests, 0 failures

[davidkoski]

Ah wait a sec -- these changes can't go here, they would need to be in mlx itself (Cmlx is mostly a submodule or files built from the submodule when I pick up a new tag).

[davidkoski]

But: Here is a comment on adding it to mlx:

ml-explore/mlx#3328 (comment)

[davidkoski]

See also #405

[RNT56]

Thanks, that makes sense. I’ll stop treating #412 as mergeable in this form.

I’ll split this up:

• withdraw/replace the Cmlx/MLX-kernel parts here, since those should not live in mlx-swift
• keep any mlx-swift follow-up limited to packaging/resource plumbing only, if that is useful without changing generated/submodule-owned Cmlx code
• keep the KV-cache integration work in mlx-swift-lm decoupled from this PR
• treat the TurboQuant kernel path as either a custom extension/custom kernel path, or something that waits for the generic quantized SDPA direction in mlx

I’ll update/close this PR accordingly.

[RNT56]

Follow-up: I split the lower-level SwiftPM metallib loader issue into ml-explore/mlx#3562, since that part belongs in MLX rather than this repo.