Stop Elixir's CLI consuming app arguments and halting the VM

[joshrotenberg]

Fixes #229.

1.6.0 split "-s elixir start_cli" from a single argv element into three. erl does not split argv elements, so the flag was never recognised before and start_cli never ran. Now it does, and it does what it is supposed to: claims --version and --help, treats the first remaining plain argument as a script path, and halts the VM when it finishes.

Two consequences for wrapped applications:

$ probe hello world
No file named hello

$ probe --version
Erlang/OTP 29 [erts-17.0.3] ...
Elixir 1.20.2 (compiled with Erlang/OTP 29)

and, with no arguments at all, an application that never halts exits about 1.3s after boot, which breaks servers and Phoenix applications.

Change

Emit --no-halt and -- after -extra, and strip them back off in Burrito.Util.Args.

--no-halt restores the pre-1.6.0 lifetime, -- stops the CLI claiming arguments, and start_cli still runs, so whatever the split in #225 was made for is unaffected.

Reverting to the single joined element would also fix both, but only by making start_cli silently not run again.

Verification

A 12-line application whose only dependency is burrito, built on macOS 26 arm64 with Zig 0.16.0, and on Linux arm64 in a container.

Arguments, before and after:

                        1.6.0                      patched
probe hello world       No file named hello        argv=["hello", "world"]
probe --version         Elixir 1.20.2 ...          argv=["--version"]
probe --help            Standalone options ...     argv=["--help"]
probe --transport stdio No file named --transport  argv=["--transport", "stdio"]

Lifetime, no arguments, application never halts:

	result
1.6.0	exits after 1289ms, rc=0
patched	still running at 12s, killed by timeout

Not broken by the change:

• exit codes propagate (System.halt(7) gives 7)
• maintenance directory and maintenance meta work
• the EPIPE handling from feat: Zig 0.16.0 compatibility #225 still terminates cleanly when piped into head -3, on both platforms
• an application passing its own -- still receives it, since only the exact leading --no-halt -- pair is stripped

Also exercised against a real application: an MCP server that takes --transport and --port and then runs indefinitely. On 1.6.0 it cannot start; with this change it parses its flags, completes a protocol handshake over stdio, serves a live tool call, and exits 0 when its client disconnects.

One note for the docs: System.argv/0 is populated only once start_cli runs, which is after Application.start/2, so applications still need Burrito.Util.Args.

[gilbertwong96]

@doawoo This PR looks good to me. Could you please merge it and bump a new version ?