fix: apply traversal-safe path normalization to direct-fetch asset fallback

[Smeagol1907]

loadMuJoCoAssets() extracts texture/mesh file paths from task XML via
/ attributes, then loads each asset
either from a zip index or by falling back to a direct fetch.

The zip-lookup path is traversal-safe: buildZipCandidates() calls
normalizeZipPath(), which strips ".." segments before using the path as a
map key. The direct-fetch fallback, however, used the raw assetPath
straight from the regex-extracted XML attribute with no normalization at
all, before concatenating it into ${apiBase}/${apiPath} and fetching it.

I haven't been able to confirm end-to-end whether task XML is ever
attacker-influenced in the production task-authoring flow (the task
generation tooling in this org looks server-side/offline from what's
public), so I can't say for certain this is reachable by an external
attacker today. Still, it's an inconsistency worth closing: one code path
in the same function treats this exact string as untrusted and sanitizes
it, the other doesn't.

Fix: normalize assetPath through the same normalizeZipPath() function
before it's used for either the zip lookup or the direct-fetch fallback,
so both paths get identical treatment. Verified normal paths pass through
unchanged and "../" sequences get safely collapsed rather than escaping
the intended directory.