Fix Numeric compatibility: hash contract, unit-preserving rounding, finite?/infinite?, to_i/to_f, and more

[nertzy]

Summary

This PR audits the full Numeric API surface against Unit's behaviour,
fixes 11 bugs found in that audit, adds test coverage for previously untested
inherited methods, adds a non-raising query API, and adds a compatibility
survey document.

Bugs fixed

#hash — violated Ruby's eql?/hash contract. Fixed as
[self.class, value, unit].hash; self.class prevents hash bucket
collisions with plain arrays of the same shape.

#ceil, #floor, #truncate — fell through to Numeric's defaults
which call to_i internally, stripping the unit. #round was already
overridden but its explicit precision arg swallowed the half: keyword.
All four now use explicit args and preserve the unit.

#finite? / #infinite? — ignored the value.
Unit(Float::INFINITY, 'm').infinite? returned nil. Both now delegate to
value.

#positive? / #negative? — raised ArgumentError for dimensional
units because Numeric's implementation calls self <=> 0. Both now
delegate to value.

#to_r — was not defined, so #numerator and #denominator raised
NoMethodError at runtime despite respond_to? returning true.
Following the Complex precedent, to_r now raises RangeError for
dimensional units and converts normally for dimensionless ones — giving
numerator and denominator for free via Numeric's implementation
(self.to_r.numerator etc.).

#remainder — raised as a downstream consequence of the #truncate
stripping bug. Fixed with an explicit implementation:
self - (self / other).truncate * other.

#to_i / #to_f — silently stripped the unit from dimensional quantities.
Following Complex#to_f's precedent (raises RangeError when imaginary part
is non-zero), these now raise for dimensional units. Dimensionless units
convert normally. Two internal callers updated: approx uses @value.to_f
directly; fdiv is overridden (see below).

#fdiv — Numeric#fdiv called self.to_f / other, which (a) now raises
for dimensional units and (b) was already returning wrong dimensions
(Unit(3,'m').fdiv(Unit(2,'m')) → Unit(1.5 m⁻¹) instead of Unit(1.5)).
Overridden as (self / other).approx — preserves correct dimension.

#real?, #real, #imag, #conj, #rect — Numeric's defaults
unconditionally treat self as real (real? always true, imag always
0, etc.), which is wrong when the scalar value is a Complex. All five now
delegate to value. This makes complex-valued units (Unit(Complex(3,4), 'ohm'))
a fully supported use case — electrical impedance, phasors, and similar
quantities are dimensional values with complex scalars.

New class methods: non-raising query API

The only way to ask "is this a valid unit string?" was to rescue two exception
classes at the call site (TypeError from validate_unit,
Unit::ParseError from parse_unit). Two new class methods handle this:

Unit.try_parse('m/s')      # => Unit("1 m/s")
Unit.try_parse('flurble')  # => nil
Unit.valid?('km/h')        # => true
Unit.valid?('m)')          # => false

Both accept an optional system argument (defaults to Unit.default_system).

New test coverage

• #hash contract, Hash key usage, Set membership
• #round with half: keyword; #ceil, #floor, #truncate with and without ndigits
• #finite?, #infinite? including Float::INFINITY cases
• #positive?, #negative? on dimensional and dimensionless units
• #to_r, #numerator, #denominator — dimensionless and dimensional RangeError
• #to_i, #to_f — dimensionless and dimensional RangeError
• #remainder including negative dividend/divisor combinations
• #fdiv — scalar divisor, compatible Unit divisor (dimensionless result), incompatible Unit divisor (combined dimension)
• #real?, #real, #imag, #conj, #rect — both real-valued and complex-valued units
• Unit.try_parse and Unit.valid?
• Multi-symbol unit aliases (sym: [a, b, ...]): all spellings resolve, first is canonical, prefix composition works on all
• +@, #abs2, #nonzero?, #integer?, #modulo, #divmod, #div, #to_int, #between?, #clamp

Survey document

doc/numeric-compatibility.md documents the full audit: which methods Unit
overrides, which inherited methods work correctly, known remaining issues
(magnitude/abs asymmetry, step raising, angle/polar for
complex-valued dimensional units), and Ruby version compatibility notes
(3.3 / 3.4 / 4.0 — the method surface is identical but Integer#** /
Rational#** overflow behaviour changed in 3.4).

Test plan

• bundle exec rake — 242 examples, 0 failures (both DSL and no-DSL runs)
• Ruby 3.3, 3.4, and 4.0