Enable link time optimization

[mngr777]

Enabling link time optimization for h3-pg/h3 improves performance by ~12--15%.

Test query:

WITH coords AS (SELECT * from generate_series(-85, 85, 0.1) AS lat, generate_series(-180, 180, 0.2) AS lon)
SELECT h3_to_geo_boundary(h3_geo_to_h3(point(lon, lat), 8))

Added compilation flags:

• -flto enables LTO;
• -fvisibility=hidden allows to inline H3 library functions in h3-pg (not allowed with default visibility since definitions can change at link time in this case);
• -fwrapv allows to inline functions like degsToRads/radsToDegs (PGXS compiles extension with -fwrapv);
• -mpc64 mitigates effect of extened floating point precision on x86 (using x87) and also improves performance by ~2--5%

Without -mpc64, LTO (inlining specifically?) causes subtle differences in floating point calculation results on x86, several tests failing for both h3-pg and h3 library. -fexcess-presicion=standard and -ffloat-store do not solve this problem completely.
Another possible solution would be to use SSE with mfpmath=sse -msse2.

[Komzpa]

Looks good to me.

[zachasme]

Looks great! Does this finally solve #23?

And how does it relate to uber/h3#360? Should any of these be part of the Release flag upstream?

[mngr777]

@zachasme
Yes, this solves #23, at least $ objdump -D h3.so | grep -E '(_square|radsToDegs)' shows no matches.
Full list of functions that are inlined or not possible to inline (with some explanation) can be obtailed with CUSTOM_COPT= <...> -fopt-info-inline-optimized-missed=inline.txt: inline.txt

The flags are related to LTO or specific to gcc, nothing to use upstream.

[zachasme]

Thanks, @mngr777! I'll merge this, and release it as part of v4 (should be in a few days).