Stop promoting double and round discrete quantiles up by default

[NAThompson]

Two breaking default changes, each restorable with one macro:

• promote_double now defaults to false on every platform, so double is evaluated in double; define BOOST_MATH_PROMOTE_DOUBLE_POLICY true to get the old behavior. Builds on @rdoeffinger's Update tests to take promote_double policy into account #1214 and Change BOOST_MATH_PROMOTE_DOUBLE_POLICY to false for non-x86 #1220. Promotion was hiding a polygamma underflow bug (fixed on develop in Keep ibeta monotonic under -ffp-contract=fast and fix develop's CI #1561); the remaining newly visible test errors (Legendre near zeros, Bessel Y'/I' near negative orders, 1F1 near b = 0, nc_t at tiny non-centrality) are all within the condition number times eps.
• Discrete quantiles now default to integer_round_up, the definition R, SciPy, Julia and Mathematica use (they agree on all 264 cases tested; Boost's old integer_round_outwards matched about half). Geometric and hypergeometric now honor the policy at exact cdf steps, and Bernoulli's complement quantile is fixed. Define BOOST_MATH_DISCRETE_QUANTILE_POLICY integer_round_outwards to get the old behavior.
• discrete_legendre (lanczos_smoothing) now uses explicit fma wherever a product meets a sum. Contraction could fuse one inlined copy of the recurrence and not another, so dlp(x, k) and dlp.next() disagreed by up to hundreds of ulps on MSVC and clang-18. Develop's CI fails the same way.

Closes #1211
Closes #1203

🤖 Generated with Claude Code

[NAThompson]

@jzmaddock , @mborland , @ckormanyos : This is the "big one", as they say. Hopefully the ulps triage merge will make this painless for everyone.