[C++][Compute] Tighten coalesce exact dispatch for parameterized varargs

[zanmato1984]

Summary

• tighten coalesce exact dispatch for parameterized decimal varargs so mixed decimal precisions/scales fall through to DispatchBest during expression binding
• keep decimal rescaling in the existing DispatchBest path
• add direct dispatch and expression-binding regressions for mixed decimal coalesce inputs

Root cause

Expression binding tries DispatchExact before DispatchBest. coalesce registered decimal varargs kernels with a broad InputType(Type::DECIMAL128/256) signature, so mixed parameterized decimals like decimal128(3,2) and decimal128(4,3) matched exactly during binding. That bypassed the decimal normalization in DispatchBest, leaving the bound expression without the implicit cast needed to a common decimal type.

What changed

• add a decimal-only exact-match constraint for coalesce varargs kernels so exact dispatch requires identical parameterized decimal types
• apply that constraint to the decimal128 and decimal256 coalesce kernel registrations only
• add regression coverage for both DispatchBest/DispatchExact and expression binding

Testing

• debug/arrow-compute-expression-test --gtest_filter=Expression.BindWithImplicitCastsForCoalesceOnDecimal
• debug/arrow-compute-scalar-if-else-test --gtest_filter=TestCoalesce.DispatchBest:TestCoalesce.DispatchExact