Fixed and UFixed models were missing minValue()/maxValue() methods,
causing the C++ range check in Block::setFixed() to be skipped.
Out-of-range values were silently clipped (positive overflow) or
stored incorrectly (negative overflow) with no error reported.

Add minValue()/maxValue() to Fixed and UFixed Python models so the
existing range check fires. Replace the silent positive-edge-case
clipping in C++ with an explicit overflow error that reports the
variable name, attempted value, and valid range.

- Use 1ULL shifts in Block::setFixed to avoid UB at 64-bit widths
- Fix UFixed.maxValue() to return the signed-range max that C++ actually
  supports (the getFixed/setFixed path is always two's-complement, so
  values above 2^(N-1)-1 are not representable for UFixed)
- Update UFixed.maximum test assertion from 4095.9375 to 2047.9375
- Replace try/except with pytest.raises(Exception, match=...) for
  stricter, idiom-consistent overflow tests

When valueBits_ == 64, casting 1ULL<<63 to int64_t gives INT64_MIN and
negating it is undefined behavior. Derive minInt from maxInt instead.

UFixed previously shared rim::Fixed's modelId and was routed through the
signed two's-complement code in Block::setFixed/getFixed, so its actual
range was the signed Fixed range rather than the [0, 2^bitSize-1]/2^binPoint
range documented in fixed_point_models.rst.

Add a dedicated rim::UFixed modelId (0x09) with Block::setUFixed/getUFixed
that enforces the unsigned range, rejects negative writes, and reads back
without sign extension. setFixedPy/getFixedPy now dispatch on modelId so the
numpy/list/scalar paths work for both. UFixed.modelId and UFixed.maxValue()
are updated to reflect the real unsigned range, and tests cover the new
full-range UFixed behavior.

Call Block::getFixed and Block::getUFixed directly in the respective
cases instead of relying on the shared getFixed_ function pointer.
The fall-through worked because getFixed_ was initialized to
&Block::getUFixed for UFixed variables, but the code was misleading
to readers.

The read path used plain-int shifts (1 << valueBits_ and
1 << (valueBits_ - 1)), which is undefined behavior for
valueBits_ >= 31 and silently broke negative-value round-trips on
Fixed variables wider than 31 bits (e.g. Fixed(32, 0), Fixed(64, 0)).

Use 1LL shifts and skip the subtraction entirely when valueBits_ == 64
(the stored value already sits in int64_t correctly). Mirrors the
write-side UB fix in 21bb136.

Add a wide-Fixed round-trip regression test covering Fixed(32, 0),
Fixed(40, 8), and Fixed(64, 0), including negative and boundary values.
CI previously only round-tripped Fixed(16, 4) / UFixed(16, 4), so the
sign-extension path was unexercised.

Update fixed_point_models.rst to reflect the split rim.Fixed /
rim.UFixed modelIds introduced earlier in this PR.