fix: prevent PHP scalar values from over-promoting array dtype - #33

This PR fixes scalar dtype promotion so that a plain PHP scalar does not force up-widening of the array's dtype — matching the intuitive expectation that $float32Array->add(5.0) stays Float32.

Motivation and Context

Previously, DType::promote() was used for both array×array and array×scalar operations. Since PHP's DType::fromValue(5.0) always returns Float64, a Float32 array combined with a PHP float scalar would promote to Float64, silently doubling memory usage and precision. Similarly, a UInt8 array plus an Int scalar would promote to Int64. This was inconsistent with how scalars combine with arrays in other numerical libraries where plain scalar values don't force up-widening.

What's Changed

• Added DType::promote_scalar() — scalar promotion that keeps float array width and integral array type, except where conversion is required (int+float → float, UInt64+signed → Float64 for overflow safety)
• Updated scalar_op_arithmetic and scalar_op_comparison macros to use promote_scalar
• Filled in missing scalar dtype conversions in get_scalar_as_u8/u16/u32/u64
• Added 6 regression tests in TypePromotionTest covering Float32+5.0, UInt8+5, Int32+10, UInt64+-1

Breaking Changes

Any code relying on scalar operations producing Float64 from Float32 arrays will now get Float32. This is a narrowing change that could affect downstream precision-sensitive code.