Fix code_cache = "off" crash in cqueues server by initializing module_reset lazily

[39george]

Summary

When running Lapis with the cqueues server and code_cache = "off", the server can crash on the second (and subsequent) request with errors coming from lua-http/cqueues such as:

❯ lapis server --config-module config development --trace
Listening on port 8080
Environment: development
2025-12-14T18:38:30 | Info  | server starting
2025-12-14T18:38:31 | Info  | database exists already
SQL: INSERT INTO "requests" ("phone_number", "text", "name") VALUES ('some number', 'Hello world', 'Some guy') RETURNING "id"
[201] POST /request - {  }
/usr/bin/lua5.4: /home/george/.luarocks/share/lua/5.4/lapis/cmd/cqueues.lua:94: ...eorge/.luarocks/share/lua/5.4/http/connection_common.lua:8: bad argument #1 to 'strerror' (number expected, got nil)
stack traceback:
	[C]: in function 'assert'
	/home/george/.luarocks/share/lua/5.4/lapis/cmd/cqueues.lua:94: in function </home/george/.luarocks/share/lua/5.4/lapis/cmd/cqueues.lua:85>
	(...tail calls...)
	...ge/.luarocks/share/lua/5.4/lapis/cmd/cqueues/actions.lua:31: in function 'lapis.cmd.cqueues.actions.server'
	(...tail calls...)
	...e/.luarocks/lib/luarocks/rocks-5.4/lapis/dev-1/bin/lapis:4: in main chunk
	[C]: in ?

This PR changes lapis/cmd/cqueues.moon (and the generated .lua) to initialize module_reset() lazily on the first stream, instead of eagerly at server startup.

Root cause

module_reset() captures a “keep” snapshot of package.loaded at the moment it’s created. When it’s created eagerly, the keep snapshot may not include all runtime server modules (probably http.* and cqueues.*), which are later loaded as the server begins handling requests.

With code_cache = "off", reset() then clears modules that should not be reloaded. This can cause lua-http / cqueues modules to be re-required multiple times and re-apply socket.interpose(...) repeatedly, resulting in layered wrappers and mismatched return values (e.g. why becoming nil or a string). Those values eventually hit ce.strerror(why) / seterror(..., why) and crash the process.

Reproduction

I'm providing an example project which reproduces the issue: https://github.com/39george/cqueues-reset-bug

Notes

I tried to set config parameter code_cache = "app_only", but it fails too with another error:

❯ lapis server --config-module config development --trace
Listening on port 8080
Environment: development
2025-12-14T18:47:42 | Info  | server starting
2025-12-14T18:47:42 | Info  | database exists already
onstream on http.h1_stream{connection=http.h1_connection{type="server";version=1.1};state="closed"} failed: /home/george/.luarocks/share/lua/5.4/lapis/cqueues.lua:29: assertion failed!