Fix potential unbounded file/type lookup

[AntoineGautier]

Latent on main, currently masked by eager type loading: dead-ended references warm typeStore
before cold lookup paths are ever exercised. The parse-skipping work for __ctrlFlow(enable=false)
(#601) removes that warming and reliably triggers RangeError: Maximum call stack size exceeded
while building Buildings.Templates.AirHandlersFans.Data.VAVMultiZone's parameter table
(schedule.test.ts, real-Buildings-library data). Root-caused with a cycle detector; this must be
fixed before or together with #601.

Root cause — four interacting defects

1. Store._generatePaths() (parser.ts:138-150) blindly prefixes already-qualified names.
   Resolving Buildings.Templates.AirHandlersFans.Data.PartialAirHandler from
   basePath = ...Data.VAVMultiZone generates candidates like
   Buildings.Templates.AirHandlersFans.Data.Buildings.Templates.AirHandlersFans.Data.PartialAirHandler.
2. getPathFromClassName()'s directory-trimming fallback (loader.ts:186-207) fuzzy-"succeeds" on
   those bogus candidates (the example above trims back to .../Data/package.json), so they load
   real files. File's within-path validation (parser.ts:1212-1218) doesn't catch it because
   within is a segment-prefix of the bogus name.
3. getFile() (parser.ts:1234-1239) has no construction dedupe: loader() memoizes lookups
   (keyed by the requested className string — each bogus candidate is a fresh key), but
   new File(...) and all element constructors re-run on every call.
4. ShortClass and Component resolve their type before registerPath (parser.ts:503 vs 512;
   829 vs 835), so duplicate punch-out never protects them (unlike LongClass, which registers at
   607 and punches out at 609). Every redundant File reconstruction re-runs a full candidate sweep
   per component/short class → more bogus candidates → more getFile calls. This closes the cycle:
   getFile ↔ pre-registration typeStore.get is mutual recursion with no terminator — genuinely
   unbounded, not just deep (a 16x larger V8 stack doesn't survive it).

Proposed fix (in order of leverage)

1. Root-aware _generatePaths: if the first segment of path names a top-level loaded package
   (derivable from MODELICA_JSON_PATH roots / PACKAGE_LIST[0]), return [path] only.
2. Dedupe File construction by resolved file identity (resolved json path or
   getClassNameFromJson): each real file constructed at most once per run. Makes the system robust
   to any future candidate-generation slop.
3. Move registerPath (and annotation parsing) ahead of type resolution in ShortClass and
   Component — also required by Fix __ctrlFlow.enable annotation handling #601.
4. Negative-cache the _load step in Store.get (a Set of candidate strings that failed to
   load; keep checking the store positively since it grows). Perf polish once 1–3 are in — it cannot
   bound the blow-up alone because the bogus keys are distinct.

Also add a permanent cheap getFile re-entrancy guard that throws with the active stack — turns any
regression here into a one-line diagnostic.