DisplayServer: Don't fallback to headless

[akien-mga]

Unless users requested the headless driver specifically, they expect to either
see a window, or that the process terminates if there's an error.

Currently it would fallback to headless so they'd unexpectedly get a valid headless
instance if their DisplayServer failed initializing (e.g. missing Vulkan support).

• Fixes Vulkan: After giving error about missing driver, Godot doesn’t exit #58414.
• Supersedes Avoid falling back to the headless display server #59074.

---

Depends on #62555, otherwise it's crashing on exit in the case where it couldn't initialize a DisplayServer.
With #62555, it prints an error instead (which itself is likely a bug that could be fixed independently):

$ VK_ICD_FILENAMES= godot-git
Godot Engine v4.0.alpha.custom_build.e7197cb6f - https://godotengine.org
ERROR: No surface extension found, is a driver installed?
   at: _initialize_extensions (drivers/vulkan/vulkan_context.cpp:353)
ERROR: Could not initialize Vulkan
   at: DisplayServerX11 (platform/linuxbsd/display_server_x11.cpp:4684)
ERROR: Unable to create DisplayServer, all display drivers failed.
   at: setup2 (main/main.cpp:1700)
ERROR: BUG: Unreferenced static string to 0: interface_added
   at: unref (core/string/string_name.cpp:131)

The last ERROR here is printed using the fallback.

[Calinou]

How does this PR compare to #59074?

[akien-mga]

How does this PR compare to #59074?

Ah thanks, I knew I remembered a PR that did something like this but couldn't find it again.

I think my solution is more elegant, poke @RandomShaper for cross-review.

[RandomShaper]

To do a recap on how this PR compares to mine:

1. This one assumes that headless will always correspond to the highest index. Mine checks by name, but to do that it has first to create and destroy it if turned out to the headless one, since there's no API to check the name beforehand.
2. Mine adds an abrupt exit (exit(255);) in case no display driver works. If I remember correctly, that was to avoid the engine startp ending in a crash that could cause further confusion to the user.

Regarding 1), I can't really tell which approach is better. Regarding 2), if it's still relevant for its purpose, I'd include it in whatever PR we end up merging.

[akien-mga]

For 1., this is seems to be the intended design that headless should always be last:

godot/servers/display_server.cpp

Line 793 in 20d4c66

// Headless display server is always last

But might need double checking. I could add e.g. an assert that the name of the last index DisplayServer is headless.

For 2., I had a crash but I fixed it with #62555. I think it's nicer to let Godot try to exit gracefully than doing a hard exit, so we can properly do finalize steps.

[RandomShaper]

Fair enough.

If we can somehow check under DEV_ENABLED that headless is the last one, cool. If not possible, I'm happy as well now I know it's like a design rule.

[akien-mga]

Added a couple DEV_ASSERT checks to validate that the headless DisplayServer is the last entry, and likewise for the dummy AudioDriver.