fix: close XML file on parse-error paths to avoid FILE* leak

[EylonKrause]

Description

ncclTopoGetXmlFromFile, ncclTopoGetXmlGraphFromFile, and ncclTopoDumpXmlToFile (src/graph/xml.cc) only call fclose() on the success path. Because NCCLCHECK(...) returns immediately on a non-success result, an error from xmlLoadSub() / ncclTopoDumpXmlRec() returns from the function before the fclose(), leaking the FILE* (and its underlying file descriptor).

This triggers whenever NCCL_TOPO_FILE or NCCL_GRAPH_FILE points at a file that fails to parse — e.g. a wrong version attribute, an unterminated/mismatched tag, or an over-long value/name. Each affected ncclCommInitRank then leaks one descriptor.

Fix: capture the result, fclose() unconditionally, then propagate via NCCLCHECK — mirroring the exit:-label cleanup discipline already used elsewhere in this file (e.g. ncclTopoGetXmlFromSys).

Related Issues

None.

Changes & Impact

• src/graph/xml.cc: three call sites (two read paths + the dump path). The only behavioral change is that the file handle is released on the error path. No API change; non-breaking.

Performance Impact

None — this is error-path-only resource cleanup.

Testing

• Builds clean with make src.build (no new warnings).
• Before/after repro with a malformed topology file (<system version="2"></system> via NCCL_TOPO_FILE), looping ncclCommInitRank and counting open fds pointing at that file:
  • Before: open fds to the topo file grow 1, 2, 3, 4, 5, 6 across six failed inits (leak).
  • After: stays at 0 on every iteration (closed).