Speed up geo coordinate transforms: combined sincos + fewer trig calls

[alexey-milovidov]

Speed up the geo coordinate transforms

The lat/lng <-> H3 cell transforms are dominated by transcendental
function calls. Profiling the hot paths showed two easy, safe wins:

1. Combined sine/cosine. Several routines need both sin(x) and
   cos(x) of the same angle but computed them with two independent libm
   calls. A small _sincos() helper computes them together — it lowers
   to a single sincos call where libm provides one, and falls back to
   separate sin()/cos() otherwise, so it stays portable.
2. Removing redundant trig. A couple of routines evaluated the same
   sine/cosine more than once.

Functions touched:

function	change
_geoToVec3d	one sincos for the latitude, one for the longitude
_geoAzimuthRads	sincos pairs; cos(p2->lat) was evaluated twice
_geoAzDistanceRads	sincos for p1->lat, distance and azimuth, each previously evaluated twice
_geoToHex2d	one sincos for theta

Correctness

The outputs are bit-for-bit identical to before — this only removes
redundant work and fuses paired calls (which return the same values).
All 311 existing unit tests pass unchanged.

Performance

Measured on 10M random cells, clang -O3, no LTO:

function	speedup
cellToBoundary	+14%
cellToLatLng	+9–11%
cellArea*	+8%
latLngToCell	+3–4%

cellToParent / isValidCell were already limited by their (already
optimal) bit-twiddling and are unchanged.

---

This backs a ClickHouse-side change: it lets ClickHouse keep the C H3
library while recovering the coordinate-transform speedups that motivated
the experiment in ClickHouse/ClickHouse#100272, without switching the
backend to Rust.

[alexey-milovidov]

Looks ok.