Setting tx levels within a group is not implemented yet. The code
lacks error checking, and is presently rather dreadful. But tests
pass.

strconv.Atoi(string([]byte)) causes the temporary string to escape
to the heap. This way we stay with a byte slice that does not escape.

VSLQ_Dispatch() is a powerful tool for the C API, but the restrictions
and overhead involved with the C/Go bridge make it too unwieldy to use.
Too much synchronization becomes necessary for a low-level API like
this. The cgocall overhead becomes excessive.

It seems to be difficult, perhaps impossible, to implement callbacks
without creating many objects on the Go heap. Since many objects from
the log are created in hot code, this puts excessive pressure on the
Go garbage collector.

This version uses low-level log read operations, excusively VSL_Next()
and related functions. The C/Go bridge is much smaller and simpler. It
should be possible to create the objects for log data on the Go stack.

The Cursor and Query interfaces reflect the underlying VSL interface
more faithfully. Distinct Cursors are safe for concurrent reads (as
for struct VSL_cursor from the VSL interface).

This means that grouping and VSL queries will have to re-implemented
in Go.

Use an atomically incremented uint64 as the key for global maps.
Random key and sync.Maps are no longer necessary.

No longer starting a goroutine -- users decide if they want one.

Native reads are locked per Log object, since they are not safe for
concurrency with the same native handles.

The global callback map is keyed with an atomically incremented uint64.

The VSM handle is exported as a field in struct VSM, for internal
use only.

Names and descriptions are stored per object, no need to sync on the
global map after Attach.