test engine: degenerate digit counts are accepted where both builders reject them

[zexoverz]

test/engine.go accepts two shapes that std/math/bits rejects during compilation, so a circuit can pass test.IsSolved and then fail to compile.

ToBinary guards only the negative case:

nbBits := e.FieldBitLen()
if len(n) == 1 {
    nbBits = n[0]
    if nbBits < 0 {
        panic("invalid n")
    }
}

n == 0 passes. FromBinary has no length guard at all, so a zero-length call is accepted.

Both builders route through std/math/bits, which rejects both:

FromBase needs at least 1 digit     conversion.go:38
nbDigits <= 0                       conversion.go:78

Observed divergence:

api.FromBinary()      engine: passes    r1cs: parse circuit: FromBase needs at least 1 digit
api.ToBinary(v, 0)    engine: passes    r1cs: parse circuit: nbDigits <= 0

api.FromBinary() diverges unconditionally. api.ToBinary(v, 0) diverges whenever the witness value of v is zero; for a nonzero v the engine panics on its own reconstruction check instead.

The stakes are much lower than #650, since both cases are compile-time rejections and nothing unsafe can ship. It is still a hole in the safety net for a gadget that computes a bit width and can legitimately reach zero, because the test engine will not tell it that the circuit is invalid.

The fix is two guards in test/engine.go. That is the same file as #1810, which is currently open, so I would rather not send a second PR against it while that is in review. Happy to fold this into #1810 if you prefer, or to send it separately once that lands. I have reproductions for both cases.