ARKit drivers: complete Phase-4 TODOs (confidence-based face source selection, shoulder/arm T-pose calibration)

Three production TODOs in the ARKit drivers fall back to placeholder behavior. Two are real feature gaps; one (weighted body-skeleton blending) is already tracked by #29.

## 1. Confidence-based face source selection
- `Sources/VRMMetalKit/ARKit/ARKitFaceDriver.swift:332`
- `.highestConfidence` source-selection mode falls back to *latest-active* — confidence is never consulted.
- Needs per-source confidence to be surfaced by the capture layer before this can be honored.

## 2. Shoulder/arm T-pose calibration (highest impact)
- `Sources/VRMMetalKit/ARKit/ARKitCoordinateConverter.swift:112`
- Shoulder offsets are **DISABLED** (`.leftShoulder/.rightShoulder = identity`) because the calibration capture was taken in A-pose, not T-pose.
- Re-capture reference offsets with the subject in T-pose and re-enable, so arm retargeting is correct.

## 3. (tracked elsewhere) Weighted body-skeleton blending
- `Sources/VRMMetalKit/ARKit/ARKitBodyDriver.swift:374` — `.weighted` mode falls back to latest-active.
- Already covered by #29; listed here only for cross-reference.

## Done when
- (1) and (2) implemented and the TODO comments removed
- ARKit driver tests cover confidence selection and T-pose-calibrated shoulder offsets

Start with the TODOs in Sources/VRMMetalKit/ARKit/ARKitFaceDriver.swift:332 and ARKitCoordinateConverter.swift:112, then inspect the capture layer and existing ARKit driver tests. Confirm how per-source confidence is exposed and how reference shoulder offsets are captured. Done means implementing the two requested behaviors, removing their TODO comments, and adding tests for confidence selection and T-pose-calibrated shoulder offsets.

Written by the indexing model from the issue text.