tests: avoid stale test file parts on retries

[GrahamCampbell]

getpart and getpartattr results are memoized keyed on section, part and filename (2039253). When a test is retried, singletest_preprocess regenerates the preprocessed file with fresh variable substitutions, such as a new server port, but the stale cached parts keep being returned, so retries reuse the old port and --retry cannot recover. One flaky hit of test 798 (the root cause of which I fixed in #22318) thus became five identical failures and a red job.

This PR adjusts loadtest to flush the memoized caches whenever it actually reads a file, and singletest_preprocess to force the reload past the already-loaded check. Repeated same-file calls still return early with a warm cache.

The same staleness also exists in ftpserver.pl, which holds its own copy of these caches in a long lived process and leans on the already-loaded check in every command handler. A retry landing on the same runner would keep serving the previous attempt's file, and that predates the memoization since the check also keeps the old file contents. The server now forces a reload whenever a client connects, while lookups within a connection stay warm.

[dfandrich]

Good catch! Not just the memoized output but the cached xml file would be affected, as you've identified. I need to think for a moment if this solution is sufficient in the case of parallel tests.

[dfandrich]

I don't see parallel tests affecting this. The selection of servers is done in the runner as is the variable substitution, so the effects should be contained within the runner process. This looks like a good solution to this.