[evaleev]

Fixes #444. Also removes the Gaussian-fit fallback from TennoGmEval: every regime now has a direct evaluator.

Yukawa core integrals for U < 1e-7 at small T (#444)

TennoGmEval::eval_yukawa() routed every U = zeta^2/(4 rho) below the interpolation range to the upward recursion regardless of T. The recursion amplifies rounding errors by ~(2m-1)/(2T) per step, so at small T the higher orders were garbage (up to 135 orders of magnitude at m = 36, T = 0.001).

Both eval_yukawa() and eval_slater() now use eval_small_U() for U < Umin, T <= Tmax: an expansion in powers of U about the Boys function,

G_m(T,U) = e^U [ sum_{k<=4} (-U)^k/k! F_{m-k}(T) + U^{m+1/2}/2 sum_{j<=4} Gamma(-m-j-1/2) (-TU)^j/j! ]

with the finite parts of F_n, n < 0, from the (stable) downward recursion. The truncation error is below 1e-20 for U <= 1e-7, T <= 1024 (2368-point grid against 50-digit quadrature, both operators; max abs error of the implementation 2.2e-16). U = 0 (zeta = 0) is now exact for any T (previously an assert/NaN).

No more Gaussian-fit fallback

The fallback was reached in two regimes; both now have direct evaluators:

• U + 2 sqrt(U T) > 709 (e.g. U >= 240 at small T, inside the interpolation range; and T > Tmax, where the upward recursion used to return NaN): only the closed forms of G_0, G_{-1} overflowed, since exp(x^2 - T) and erfc(x) were formed separately. They are now computed as exp(-T) erfcx(x) for x > 25 (asymptotic series of erfcx), which never overflows.
• U > Umax = 1e3: new eval_large_U(), the asymptotic (Watson) expansion of G_m = (e^{-T}/2) int_0^inf e^{-(U-T)s} (1+s)^{-m-3/2} e^{-T s^2/(1+s)} ds in 1/(U-T), with a 3-term recursion for the coefficients. G_{m-1} - G_m (Slater) has its own series, avoiding cancellation. Values are below 1e-300 unless U - T > 255, where the series reaches double precision in a few dozen terms.

Removed: the cache of Gaussian fits (which leaked by design), both GaussianGmEval members (not safe to use concurrently although the TennoGmEval instance is shared by all threads), the unconditional clog diagnostics, and the bool/optional failure plumbing. fit_slater() is kept as a public utility.

regime (mmax = 36)	before	after
U = 800, T = 0.3	Gaussian fit, 40 us	interpolation, 6.6 us
U = 1e4	Gaussian fit, 40 us	asymptotic series, 0.5 us, rel. err. <= 6e-16
T = 1280, U = 125	NaN	1.7585902911277e-295 (exact: 1.7585902911279e-295)

Tests

test-core-ints.cc:

• small U: Yukawa at T = 1e-3, 1, 1024 and Slater at T = 1e-3, 1 with U = 1e-8; U = 0 against FmEval_Chebyshev7; T = 0, U = 1e-8 tolerances tightened from 1.4e-11 to 1.4e-14;
• former fallback regimes, both operators: T = 0.3, U = 800 and T = 5, U = 1000 at the table's absolute precision; T = 0.3, U = 1e4, T = 10, U = 2000, T = 0, U = 1e6 at 1e-14 relative precision (40-digit references); T > Tmax overflow cases finite and negligible;
• "TennoGmEval shared across threads": every path from 4 threads, bitwise agreement with the serial result (failed under ThreadSanitizer with the old fallback; clean now).