Apply the portrayal script restriction to package.loaded as well

[SemanticWave-Hoyeon]

What this changes

Portrayal catalogues arrive with an exchange set and are executed by this interpreter, so the constructor already tried to keep the standard library away from them by assigning nil to a few globals.

That does not reach far enough. luaL_openlibs() also records every module it opens in package.loaded, and that table was left untouched, so require() returned the same module the global had just been cleared of. Clearing the global changes what the name resolves to; it does not remove the module.

The same applies to the ffi module on the LuaJIT runtime this project links against (extlibs/Lua is LuaJIT 2.1.0-beta3, despite what readme.txt says), and that module was not covered by the original list at all.

The change

The restriction now clears each module from _G, package.loaded and package.preload, drops the native module loaders together with package.cpath and package.loadlib, and removes the chunk loading entry points.

Two things are deliberately left alone:

• pure-Lua require() keeps working, because rule files rely on it to pull in sibling rules (require 'PortrayalModel', require 'RESTRN01', and so on). Only the native loaders are removed, so module resolution still finds .lua files and no longer finds anything else.
• rawget stays, because the rule files use it heavily (35 call sites across the shipped catalogues). rawset is cleared, as before, and is unused by them.

The result of the chunk was previously discarded. A restriction that failed to apply would have gone unnoticed while the rest of the code kept assuming it held, which is the part that seemed most worth fixing. It is now checked, and a second chunk verifies that the modules and loaders are genuinely unreachable before any catalogue runs. Either failure closes the state and aborts construction rather than continuing unrestricted.

That last point is a behaviour change worth flagging: lua_session's constructor can now throw. It only does so if the restriction did not take effect, which should not happen, but failing closed seemed the right default for something whose whole job is to hold. Say the word if you would rather it logged and continued.

Verification

Checked against LuaJIT 2.1, matching the vendored runtime. Before the change, require('os'), package.loaded.os, require('io'), require('debug') and require('ffi') all still resolve after the existing block runs. After it, each of those is refused, the chunk loaders are gone, and the verification chunk passes. A require of a plain .lua module still loads normally, and rawget is intact.

No .lua file is touched by this PR, and nothing under extlibs/ is modified. As with the companion parser PR, I could not build the full MSVC/MFC project here, so this has not been compiled in place or run through the application.