Read the REAL mantissa as an unsigned integer and check exponent scaling

[dzatona]

Addresses the first and third defects in #143. The exponent-format defect is
left alone — that one changes which inputs are accepted, so it is a separate
decision.

Two changes in decode_real:

The mantissa is accumulated into u64 instead of i64, with the sign applied
at the conversion to f64. X.690 8.5.7.5 defines N as an unsigned binary
number, so bit 63 of an eight-octet mantissa is a value bit, not a sign bit.
Today 09 0a 80 00 80 00 00 00 00 00 00 01 decodes as
-9.223372036854776e18 in every profile; with this it decodes positive. The same
change removes the negation overflow, since there is no longer a negation of
i64::MIN to overflow.

u64 rather than i128 because the existing rem.len() > 8 guard already caps
N at eight octets, so u64 represents every mantissa the crate accepts, exactly.

checked_mul on the base-8 and base-16 exponent scaling, returning
InvalidValue { msg: "Exponent out of range (REAL)" }. This is reachable on
BER-conforming input today — a 37-byte value with X = 32 gets its exponent
misread as 20 01 00 00, and e * 4 leaves i32 before the mantissa-length
guard is consulted — and it stays reachable if the exponent-format defect is
fixed, because X = 4 remains legal and a four-octet exponent spans i32. So
this part stands on its own regardless of what happens to the format handling.

Three tests in tests/ber.rs, in the style of the existing from_ber_real_*
ones: the unsigned mantissa, the most-negative mantissa, and four inputs that
must be rejected for exponent range — the two short malformed ones plus both
minimal BER-conforming forms, 48 bytes for base 8 and 37 for base 16. All three
fail against an unpatched real.rs.

cargo fmt --check, cargo clippy --all-features --all-targets -- -D warnings
and cargo test --all-features are clean: 466 tests, no existing test changed.
Both build profiles now agree on every input I have, which they did not before.

The CI failures on this branch are the compile_fail trybuild mismatch already
present on master — a rustc diagnostic reworded since the .stderr was
recorded. Not related to this change.