BUG: inter-backend casts lose precision by converting through double

Conversions between the SLEEF and long-double backends currently use double as an intermediate in quad_to_quad_strided_loop.

Examples from casts.cpp:

Sleef_cast_from_doubleq1(static_cast<double>(ld));
static_cast<long double>(cast_sleef_to_double(sleef));

This discards valid precision beyond binary64 in both directions. In particular, x87 long double has a 64-bit significand, while binary64 has only 53 bits.

A value such as 1 + 2**-60 is exactly representable by x87 long double and binary128, but is rounded to 1.0 by the current long-double -> SLEEF cast.

The casts should convert directly between the source and destination formats, preserving signed zero, finite values, subnormals, infinities, and signed NaNs. SLEEF -> long double may be narrowing, but it should round directly to the target format rather than first narrowing to double.

Acceptance criteria:

• Long-double -> SLEEF preserves every long-double value exactly.
• SLEEF -> long double is correctly rounded for the platform format.
• No conversion passes through binary64.
• Boundary, subnormal, signed-zero, infinity, NaN, and precision-sensitive round-trip tests are included.