Compare floats tolerantly in round-trip test instead of ignoring them

[argonui]

Problem

Fixes #98

The end-to-end round-trip test tests/e2e_test.go treated every float64 as unpredictable and skipped it entirely:

if _, ok := v.(float64); ok {
    return true
}

That meant numeric corruption in the reverse→build round trip passed undetected — the very thing the round-trip property is supposed to guarantee.

Fix

Stop skipping floats. Compare them approximately with cmpopts.EquateApprox(0, 1e-4). Only Date and EpochTime remain ignored, since they are regenerated on every build by design.

The near-identical ignoreUnpredictable in root compare_test.go got the same treatment (a shared approxFloats option applied to both cmp.Diff calls).

Tolerance choice: 1e-4

Number smoothing (see objects/AGENTS.md) rounds positions to 3 decimal places, scale to 2, and colors to 5. The coarsest of those, 3dp, means legitimate values differ by at most 5e-4 from an unrounded input. An absolute tolerance of 1e-4 sits comfortably below any real change a user could make in-game (nudging a card moves it far more than 1e-4) while staying above floating-point rounding noise. It is tight enough to catch real numeric corruption yet loose enough not to flag representation jitter.

Fixture normalization

For a tolerance this tight to hold, the compared floats must be in their committed (smoothed) form — the form the tool always emits. 3_snap_points.json already stores whole-degree rotations, but two fixtures (GHE_Dev_no_objects, GHE_Dev_no_objects_no_lua) still contained raw TTS snap-point Position/Rotation values (e.g. 180.0298, -36.0014) that the tool smooths on every build and therefore never round-trip identically. Those snap-point values are smoothed in place, exactly matching the tool's own output (verified programmatically). No other fields changed. Rotations round to whole degrees, which is why raw snap-point rotations could not survive a tight tolerance untouched.

Proof the test now catches corruption

Temporarily mutated smoothPos in objects/numbersmoother.go to roundFloat(f, 3) * 2:

• Before this change (blanket float skip): go test ./tests/ passed — corruption invisible.
• After this change, with the same mutation: go test ./tests/ FAILS. The fixtures carrying position data (3_snap_points, GHE_Dev_no_objects, GHE_Dev_no_objects_no_lua) report the doubled coordinates; position-free fixtures still pass.

Reverting the mutation restores full green: go test ./..., go build ./..., and go vet ./... all pass.