[BACKEND][CPU] Convert tt.get_program_id and tt.print (Hello World)

[minjang]

Summary: tl.program_id needs to be lowered first for any meaningful example. As of now, we think pid will be provided as additional function arguments to the kernel. The CPU launch (parallel for) will give grid ids. Getting program_id is mapped to reading one of the last three arguments.

I also quickly implemented tl.device_print or print, only for scalar types for a quick "Hello, World!" testing.

Test Plan: Tested with a simple example:

@triton.jit
def add_kernel(...):
    pid = tl.program_id(axis=0)  # We use a 1D launch grid so axis is 0.
    foo = pid + 42
    tl.device_print("Hello, World!", foo, pid)

The resulting .llir is valid:

@printfFormat_1 = internal constant [31 x i8] c"pid (%u, %u, %u) test: %u, %u\0A\00"

declare !dbg !3 i32 @printf(ptr, ...)

define void @add_kernel(ptr addrspace(1) %0, ptr addrspace(1) %1, ptr addrspace(1) %2, i32 %3, i32 %4, i32 %5, i32 %6) !dbg !7 {
  %8 = add i32 %4, 42, !dbg !8
  %9 = call i32 (ptr, ...) @printf(ptr @printfFormat_0, i32 %4, i32 %5, i32 %6, i32 %8, i32 %4)
  ret void, !dbg !9
}

Tried to compile with a fake main function:

> % cat main.c
extern void add_kernel(float*, float*, float*, int, int, int, int);

int main() {
    add_kernel(0, 0, 0, 4, 5, 6, 7);
}

> % llc -filetype=obj add_kernel.llir && clang -o a.out add_kernel.llir.o main.c
> % ./a.out
pid (5, 6, 7) Hello, World!: 47, 5