Inconsistency in the way promotion to `long double` is handled between x86_64 and other architectures, e.g. aarch64

[AWSjswinney]

Description

I want to propose disabling promotion to from double to long double by default in Boost's math functions due to performance implications and architectural differences between platforms. Users who need higher precision can enable promotion with a policy.

Boost documentation on type promotion: https://live.boost.org/doc/libs/1_44_0/libs/math/doc/sf_and_dist/html/math_toolkit/policy/pol_ref/internal_promotion.html

Current Behavior

• On x86_64: Promotes to 80-bit extended precision long double (15-bit exponent, 63-bit precision)
• On aarch64: Promotes to 128-bit quadruple precision long double (15-bit exponent, 112-bit precision). The ARMv8 ISA does not define a floating point type wider than 64 bits and so GCC uses a soft float implementation for the quadruple precision, 128-bit float.

Issues with Current Behavior

1. Performance Impact

   • Significant performance overhead from promotion to long double
   • Larger performance delta between architectures (e.g., x86_64 vs aarch64) compared to standard library implementations

2. Architectural Inconsistency

   • Different precision levels between platforms (63 bits vs 112 bits)
   • Potentially unexpected behavior when moving code between architectures

3. Limited Benefit

   • Numerical differences at these precision levels are typically negligible for most use cases
   • Most applications that truly need extended precision would explicitly specify it