Add HELION_PRINT_REPRO=1 to print Helion kernel repro script to console

[yf225]

Our users commonly invoke Helion kernel from a very deep code stack, and it's been very difficult to extract a small repro.

With HELION_PRINT_REPRO=1, we will print the minimal Helion kernel repro script to console, thus making it much easier to repro issues.

Example output to stderr:

import helion
import helion.language as hl
import torch
from torch._dynamo.testing import rand_strided

@helion.kernel(config=helion.Config(block_sizes=[2, 2], flatten_loops=[False], indexing=['pointer', 'pointer'], l2_groupings=[1], load_eviction_policies=[''], loop_orders=[[0, 1]], num_stages=1, num_warps=4, pid_type='flat', range_flattens=[None], range_multi_buffers=[None], range_num_stages=[0], range_unroll_factors=[0], range_warp_specializes=[None]), static_shapes=True)
def kernel1(x: torch.Tensor) -> torch.Tensor:
    out = torch.empty_like(x)
    m, n = x.shape
    for tile_m, tile_n in hl.tile([m, n]):
        out[tile_m, tile_n] = x[tile_m, tile_n] + 1
    return out

def helion_repro_caller():
    torch.manual_seed(0)
    x = rand_strided((2, 2), (2, 1), dtype=torch.float32, device=DEVICE)
    return kernel1(x)

[oulgen]

i dont think you need any of this, i added a mode to to_triton_code that prints repro?

there's even a unit test for it

[yf225]

yeah so I actually want to get the Helion kernel (with config and caller) not the generated Triton kernel.

what internal folks usually provide is a buck command that's many layers deep, and it's very hard to extract out a clean Helion kernel repro. With this env var, it will print the Helion kernel with the right input tensors, so that I can just focus on debugging the minimal repro.