improved gcc-9 and gcc-10 decoding speed

[Cyan4973]

I was embarrassed that the updated decoder code
would fare measurably worse on gcc-9 and gcc-10
by 5 to 6%.

The new alignment setting is better for gcc-9 and gcc-10 by almost +5%.
Consequently, this version still decodes slower than v1.4.9 on gcc-9 and gcc-10,
but by a more negligible amount < 2%.

Unfortunately, the new setting is worse for essentially all other compilers.
Make the new alignment setting conditional to gcc-9+,
keep previous alignment setting for all other compilers.

align6+5+(3-4)
Compiler	silesia-L1	enwik8-L1	silesia-L5	enwik8-L5	enwik9-L22	enwik9-long30
gcc-7	1671	1519	1459	1205	1033	1029
gcc-8	1604	1418	1409	1152	1013	1010
gcc-9	1630	1442	1434	1176	973	971
gcc-10	1650	1456	1448	1185	919	917
clang-6.0	1543	1341	1379	1115	1037	1033
clang-7	1495	1334	1330	1098	1015	1013
clang-8	1536	1356	1317	1049	995	994
clang-9	1585	1403	1395	1131	1039	1033
clang-10	1590	1404	1387	1114	1019	1015
clang-11	1515	1337	1347	1103	1007	1004

Comparison with dev
Compiler	silesia-L1	enwik8-L1	silesia-L5	enwik8-L5	enwik9-L22	enwik9-long30
gcc-7	-0.06%	-0.07%	-0.07%	-0.41%	-0.10%	-0.29%
gcc-8	-0.12%	-0.21%	-0.14%	-0.60%	-0.10%	-0.10%
gcc-9	2.64%	1.91%	5.44%	6.33%	-0.10%	0.00%
gcc-10	2.87%	2.32%	5.62%	5.99%	0.11%	0.00%
clang-6.0	0.06%	-0.45%	-0.22%	0.00%	0.10%	-0.10%
clang-7	-0.13%	-0.07%	0.00%	-0.18%	-2.96%	-2.50%
clang-8	-0.19%	-0.59%	0.00%	-0.29%	-1.29%	-1.09%
clang-9	-0.13%	-0.28%	-0.07%	-0.09%	0.29%	0.00%
clang-10	0.00%	-0.35%	-0.86%	-1.68%	-0.39%	-0.59%
clang-11	-0.07%	-0.15%	0.15%	0.36%	0.30%	0.10%
average	0.49%	0.21%	0.98%	0.94%	-0.41%	-0.46%

As expected, the change is mostly neutral for all other compilers
(one can safely ignore anything <1% as "noise"),
only gcc-9 and gcc-10 in "normal" mode are favorably impacted, and by a nice margin.