fix: preserve numeric comparison precision

[kaptinlin]

Numeric comparisons currently convert integer operands to int, so uint64(math.MaxUint64) > uint64(0) evaluates to false and uint64(math.MaxUint64) == int64(-1) evaluates to true. Converting integers to float64 also makes distinct values beyond 2^53 compare equal.

Preserve signedness and width in the comparison generator. Mixed integer/float comparisons check the integer range and compare the integral part before the fractional part, preserving IEEE NaN ordering. Regenerate the five comparison helpers and add public Compile/Run regressions for typed and dynamic environments.

Validation on Go 1.27:

• go test -p 1 ./...
• go test -race -p 1 .
• go test -p 1 -tags=expr_debug -run=TestDebugger ./vm
• go vet ./...
• go test -p 1 -ldflags='-s -w' -run '^$' -fuzz '^FuzzNumericComparison$' -fuzztime=20s -parallel=1 . passed 392,471 inputs against a math/big.Rat oracle.
• Formatting, regeneration consistency and clean-base patch application passed.

A temporary go-valueexpr workspace also reproduces the corrected comparisons while preserving uint64 identity and its existing Shape type checks. Arithmetic helpers are byte-identical to the base.