Fix bad_alloc in verbose depth-map point-cloud export

[Rheinwalt]

Fix a crash when running DensifyPointCloud with high-resolution depth maps and verbose logging enabled.

At -v 3, dense reconstruction exports per-depth-map debug PLY files. The preallocated PLY memory buffer size can overflow for large depth maps, producing a huge allocation request and aborting with std::bad_alloc.

This avoids the preallocated memory buffer for these debug point-cloud exports and uses the normal streamed PLY write path instead.

[cdcseacave]

thank you for spotting the bug, but PLY writing with buffer is much faster, instead of removing it can you fix the buffer size approximation pls?

[Rheinwalt]

Maybe you're right, no buffer was the lazy way out. However, with many threads and large images this can be pretty memory intensive still.