Validate stereo fusion output path before running fusion

[behnamasadi]

RunStereoFuserImpl validated workspace_format, input_type and output_type up front, but never checked output_path until the results were written by WriteBinaryPlyPoints / Reconstruction::Write* after fusion completed. On large datasets fusion can run for hours, so a misspecified output_path (e.g. an existing directory when output_type is PLY, or a non-existent parent directory) aborted the whole run at the very end with a deferred THROW_CHECK_FILE_OPEN, discarding all the fusion work (issue #3117).

Add an early, output-type-aware validation of output_path in RunStereoFuserImpl before constructing the fuser:

• PLY writes a single file, so reject an existing directory and require the parent directory to exist.
• BIN/TXT write a directory of files, so require it to already exist (matching the THROW_CHECK_DIR_EXISTS in Reconstruction::Write*).

Also add a ply_test case asserting that WriteTextPlyPoints / WriteBinaryPlyPoints on a directory path fail cleanly, documenting the underlying deferred-failure behavior the early check guards against.

Fixes #3117

[behnamasadi]

Flagging an overlap I should have caught earlier: #4519 by @alsolex predates this PR by about a week and fixes the same issue (#3117) with essentially the same up-front validation. Two open PRs for one bug is probably part of why neither has moved, so to make the choice easy:

I've just added the THROW_CHECK_HAS_FILE_EXTENSION(output_path, ".ply") check from #4519, which mine was missing. This branch is now a strict superset of #4519:

• same up-front validation of output_path before fusion starts,
• plus the .ply extension check from Handle directory output_path for stereo_fusion PLY output #4519,
• plus a directory-vs-file distinction for the bin/txt output types,
• plus a regression test in src/colmap/util/ply_test.cc covering the directory-as-output-path case,
• plus error messages that name the offending option.

@alsolex — no ownership claim here, you got there first. I'm equally happy for #4519 to be merged and to send my test as a follow-up PR on top of it; whichever the maintainers prefer. @ahojnnes, either path closes #3117, so this just needs a pick.