Free the C string MEOS hands back, in the wrappers that convert one

[estebanzimanyi]

WITNESS: 125 generated wrappers end return C.GoString(_cret), nil and none of
them frees _cret. The generator is not free-blind -- it emits C.free 295
times, every one for an INPUT it allocated itself (_c_endian, _c_name,
_c_path). It frees what it allocates and drops what MEOS hands back.

The return belongs to the caller, and MEOS's own example says so:
meos/examples/03_berlinmod_assemble.c reads
char *trip_str = temporal_as_hexwkb(trip, WKB_EXTENDED, &length); and then
free(trip_str). C.GoString copies into Go memory, so the C allocation is
unreachable the moment such a wrapper returns.

MEASURED, 200000 calls of TemporalAsHexwkb against resident memory: the
generated wrapper grows +19444 KB where this package's hand-written accessor,
which frees, grows +4488 KB. With the free emitted the same wrapper grows
+424 KB.

THE DISCRIMINATOR IS THE C DECLARATION, so nothing is inferred from a name or a
group: char * belongs to the caller, const char * belongs to MEOS and must
never be freed. tools/codegen.py carries separate rows for the two
types whose conversions otherwise coincide; the free hangs off that distinction
and is deferred, so it runs after the copy, and it sits after the errno guard so a
failed call never reaches it.

The regenerated surface splits exactly along that line: of the 125 wrappers,
115 free and 10 do not, and those 10 are precisely the const char * entries --
geo_typename, interptype_name, meos_pc_schema_compression,
meos_pc_schema_xml, meosoper_name, meostype_name,
temporal_basetype_name, temporal_interp, temporal_subtype,
tempsubtype_name. Freeing one of those would be a fault, so the zero there
matters as much as the 115.

The suite reads 121 result lines, 0 skipped, and go build ./functions is
clean.