Fix one-sample angular velocity spike when the IMU rotation passes 2nπ

[hijimasa]

Fixes #155

Symptom

The moment an object carrying the IMU passes a multiple of 2π of
accumulated rotation (i.e. every full turn), the reported angular
velocity goes wild for exactly one sample — roughly 2π/dt, which is
about 7200 deg/s at a 50 Hz update rate.

Root cause

Quaternions double-cover rotations: q and −q describe the same
attitude. Consecutive transform.rotation samples can land on opposite
hemispheres of that cover, and when they do, the delta
Quaternion.Inverse(last) * current comes out with w < 0.
ToAngleAxis on such a quaternion reports the long way around —
~(360° − δ) about the inverted axis instead of the true small increment
δ. Divided by dt, that is the one-sample spike, and it lands exactly at
full-turn boundaries.

Fix

• When the delta quaternion has w < 0, negate all four components
  before calling ToAngleAxis (a no-op as a rotation), so the short
  arc is always measured.
• To make the math unit-testable, the delta-to-rate computation is
  extracted into public static IMUSensor.AngularVelocityBetween( previous, current, dt), called from UpdateSensorOnce(). Behaviour
  is otherwise unchanged.

Tests (EditMode: Tests/Editor/ImuAngularVelocityTests.cs)

Test	What it checks	Before fix	After fix
SmallStep_ReportsTheRotationRate	a plain small step comes out as angle / dt	Pass	Pass
FullTurnBoundary_DoesNotSpike	a 2° step across the full-turn boundary (AngleAxis(359°) → AngleAxis(1°)) — the exact scenario of this issue	Fail (reports ~358°/dt)	Pass
NegatedRepresentation_IsTheSameRotation	feeding the −q representation yields the same rate	Fail	Pass

Verified with the Unity Test Runner (6000.3.21f1, batchmode, EditMode):
3/3 pass with the fix; with the fold removed, exactly the two boundary
tests fail while the plain-step test still passes, so the tests isolate
this specific defect.

Impact

All downstream consumers of the angular velocity. Attitude-integrating
consumers (e.g. an AHRS) are barely affected, because rotating by
(2π − δ) about the inverted axis is almost the same attitude as
rotating by δ about the true axis — but anything using the raw rate
(rate logging, rate sanity checks that drop the sample, controllers)
sees a real glitch once per full turn.