Use stream channel for size lookup

`database.length()` is a path-based `stat(2)`, which can race with file
replacement between the `FileInputStream` open and the size lookup —
the bytes the stream then delivers may not match the size we read.
Switch to `stream.getChannel().size()`, which is an `fstat` on the
already-open file descriptor and therefore atomic with the open.

`FileChannelImpl.getChannel()` is lazy and `.size()` is a stat-only
syscall, so this does not reintroduce the per-thread direct ByteBuffer
cache that the previous commit was fixing — that cache only grows
when `FileChannel.read(heapByteBuffer)` is called.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: oschwald

src/main/java/com/maxmind/db/BufferHolder.java

@@ -31,7 +31,9 @@ final class BufferHolder {
             // mode that would mean chunkSize bytes of off-heap memory held per loader
             // thread for the JVM's lifetime.
             try (FileInputStream stream = new FileInputStream(database)) {
-                long size = database.length();
+                // Size from the open fd (fstat) so it's atomic with the open;
+                // getChannel().size() does not populate the buffer cache.
+                long size = stream.getChannel().size();
                 var name = database.getName();
                 if (size <= chunkSize) {
                     this.buffer = SingleBuffer.wrap(readFully(stream, (int) size, name));