Packed sub reverts ExponentUnderflow near EXPONENT_MIN where the unpacked reference path succeeds

[thedavidmeister]

Found by the testSubPacked fuzz on PR #270's CI run (seed-dependent; the PR's diff is dep pins only and does not touch this code path).

Counterexample

testSubPacked(
    Float.wrap(0x8000000000000000000000000000000000000000000000000000000000000003),
    Float.wrap(0x800000000000000000000000000000000000000000000000000000000000000a)
)

i.e. a = 3e-2147483648, b = 10e-2147483648 (both at the packed exponent minimum).

• Reference path: LibDecimalFloatImplementation.sub(3, -2147483648, 10, -2147483648) succeeds, and packLossy of its result succeeds.
• Packed path: LibDecimalFloat.sub(a, b) reverts ExponentUnderflow(-7e75, -2147483723) — the -7 difference gets normalised up to full precision (-7e75), pushing the exponent 75 below int32 min.

The test's contract is that the two paths agree (same result or same revert), so this is a genuine inconsistency at the exponent floor, not a test artefact.

Pre-existing

Replayed bit-for-bit on unbumped main (bcd9ed9, solmem 0.1.3 / datacontract 0.1.0) and on the #270 branch: identical revert, identical gas (9566) on both. Latent on main; CI's random fuzz seed on #270 happened to find it.

Repro harness used

A concrete test mirroring testSubPacked's body with the counterexample inputs wrapped via Float.wrap, run in the flake shell on both refs.