tgamma overflows with half-integer arguments

The following MRE illustrates that tgamma overflows (i.e. returns infinity) when half-integers are used:
```
using namespace boost::multiprecision;
typedef number::digits, backends::digit_base_2, void, boost::int16_t, -1022, 1023>, et_off> float8x;

long double t = 3.5l;
float8x tt = 3.5l;

long double r = tgamma(t);
long double rr = (long double) tgamma(tt);
```
Produces 3.32335097044784255117 for `r`, but infinity for `rr`. In contrast:
```
float8x tt = 3.l;
long double rr = (long double) tgamma(tt);
```
Produces the correct value of 2.

(g++ 8.3.0 on Linux, `long double` is 80bit extended)

Start by compiling the provided minimal reproduction with the cpp_bin_float configuration and compare tgamma for 3.5 and 3.0 against long double. Trace the tgamma path used for half-integer cpp_bin_float values, then verify that the 3.5 result is finite and matches the reference behavior without regressing the 3.0 case.

Written by the indexing model from the issue text.