Use llvm 13+'s [[musttail]] for dynapi

[madebr]

Suggested here as an attempt to reduce file sizes (#9206).

When building a 32-bit msys2 llvm15 toolchain, file size reductions are 10 kiB to nothing (a stripped executable is around 1.88MB)

• Because llvm's [[musttail]] requires an explicit return, this pr introduces SDL_DYNAPI_PROC for functions returning something and SDL_DYNAPI_PROC_VOID for functions returning nothing. It also introduces SDL_DYNAPI_PROC_NO_TAILCALL for explicitly disabling [[musttail]] (see below why). gendynapi.py has been modified to support these.
• Introduce __attribute__ ((musttail)) in appropriate locations in SDL_dynapi.c.

[[musttail]] has been explicitly disabled for functions accepting a SDL_GUID argument by value. Testautomation was segfaulting. I'm not entirely sure why this happens.
Ironically, disabling [[musttail]] (for DSL_GUIDToString) produces less code that works and code that uses a tail call.

SDL_GUIDToString, built with llvm's [[tailcall]] (segfaults)

1001bc80 <_SDL_GUIDToString>:
1001bc80: 8b 4c 24 18                   mov     ecx, dword ptr [esp + 0x18]
1001bc84: 8b 54 24 14                   mov     edx, dword ptr [esp + 0x14]
1001bc88: a1 e0 92 1b 10                mov     eax, dword ptr [0x101b92e0]
1001bc8d: 0f 10 44 24 04                movups  xmm0, xmmword ptr [esp + 0x4]
1001bc92: 0f 11 04 24                   movups  xmmword ptr [esp], xmm0
1001bc96: 0f 11 44 24 04                movups  xmmword ptr [esp + 0x4], xmm0
1001bc9b: 89 54 24 14                   mov     dword ptr [esp + 0x14], edx
1001bc9f: 89 4c 24 18                   mov     dword ptr [esp + 0x18], ecx
1001bca3: ff e0                         jmp     eax
1001bca5: 66 66 2e 0f 1f 84 00 00 00 00 00      nop     word ptr cs:[eax + eax]

SDL_GUIDToString, build without llvm's [[tailcall]] (works)

1001bc80 <_SDL_GUIDToString>:
1001bc80: ff 25 e0 92 1b 10             jmp     dword ptr [0x101b92e0]
1001bc86: 66 2e 0f 1f 84 00 00 00 00 00 nop     word ptr cs:[eax + eax]

FYI, llvm generates code that is peculiar:

1001bcb0 <_SDL_GamepadConnected>:
1001bcb0: 8b 44 24 04                   mov     eax, dword ptr [esp + 0x4]
1001bcb4: 89 44 24 04                   mov     dword ptr [esp + 0x4], eax
1001bcb8: ff 25 e4 92 1b 10             jmp     dword ptr [0x101b92e4]

Does anybody know why this happens? The mov's can be removed, can't they?
I observe the same behavior in other SDL dynapi entry functions accepting an argument, and it is worse for functions accepting multiple arguments.
e.g.

1001c5c0 <_SDL_GetMasksForPixelFormatEnum>:
1001c5c0: 0f 10 44 24 0c                movups  xmm0, xmmword ptr [esp + 0xc]
1001c5c5: 8b 44 24 08                   mov     eax, dword ptr [esp + 0x8]
1001c5c9: 8b 4c 24 04                   mov     ecx, dword ptr [esp + 0x4]
1001c5cd: 89 4c 24 04                   mov     dword ptr [esp + 0x4], ecx
1001c5d1: 89 44 24 08                   mov     dword ptr [esp + 0x8], eax
1001c5d5: 0f 11 44 24 0c                movups  xmmword ptr [esp + 0xc], xmm0
1001c5da: ff 25 94 94 1b 10             jmp     dword ptr [0x101b9494]

Description

Existing Issue(s)

[madebr]

The SDL_GUID issue might not be detected because of the lack of a warning:
llvm/llvm-project#56676

llvm/llvm-project#72390 is also what happens (I think)

[madebr]

I don't see the weird x86 code on Linux, when building with a (slightly) more modern llvm toolchain.

[okuoku]

musttail sorta premature on older Clang e.g.) llvm/llvm-project#50758 so it'd be safer to add some escape hatch that ignores HAVE_ATTRIBUTE_MUSTTAIL (or just filter out older clang there)