OPTIM: http: simplify http_get_status_idx() using a hash

OPTIM: http: simplify http_get_status_idx() using a hash

This function uses a large switch/case, but the problem is that due to the
numerous holes in the range, the compiler implemented a large jump table.
With a bit of experimentations, some trivial perfect-hash code works, and
since the number of entries is 19, it was enlarged to match the nearest
next power of two to avoid a large modulo operation, and fills the holes
with the default return value (HTTP_ERR_500). Jumping to 32 also results
in a lot of valid keys and allows us to pick small values, resulting in
very fast and compact code. The new function, despite keeping a 32-bytes
table, saves slightly more than 800 bytes of code+data compared to the
previous code, and avoids table jumps that affect the CPU's branch history.

Note that another simple hash worked fine and produced exactly 19 codes
(hence no need to pad holes): ((status * 8675725) >> 13) % 19

But it's still about 24 bytes larger in code to save 13 bytes of data
that are aligned anyway, and it was a bit more expensive so that was
definitely not worth it.

The validity of the table was verified with this test code added just after
it:

  __attribute__((constructor)) void http_hash_test(void)
  {
	int i;
	for (i = 0; i <= 600; i++)
		printf("code %d => %d\n", i, http_get_status_idx(i));
	exit(0);
  }

And starting haproxy |grep -vw 14 correctly shows all ordered values
(except 500 of course which is 14).

In case new codes would be added, just play again with dev/phash to
updated the table. As long as there are less than 32 effective entries
it will remain easy to update without having to modify phash.