Use the optimized multiply path for decimal32_t

[ibmibmibm]

• decimal32_t * decimal32_t gave the raw product to the constructor. The
  other types take the path through mul_finalize_u64, which is faster.
  A comment said that the hardware divide of the constructor was cheap
  enough, but on the machine below the constructor path is slower.
• decimal32_t now takes the same path as the other types, except for one
  case. That path rounds the product to 7 digits before pack_in_range,
  and the constructor then rounds a subnormal result a second time.
• Because of this, a product with an exponent below etiny still goes to
  the constructor unrounded. Above etiny, the rounded result has 7 digits
  or is at least 1e-94, so one rounding is correct.
• decimal32_t * Integer keeps the constructor path. The optimized path
  gives a zero product a different quantum exponent there.

The results are bit-identical to develop. I compared 100 million random
products in each of the five rounding modes, with subnormal operands, short
significands and integer right operands.

test/benchmarks.cpp, multiplication, median of 6 runs (Ryzen 9 3900X,
GCC 16, -O3):

Type	develop	this PR
decimal32_t	3376647 us	2650213 us
decimal_fast32_t	2006705 us	2006168 us

The checksum s of the benchmark is the same before and after.

[mborland]

Thanks!