Support for tl shift operator on tensor about triton HOT 3 OPEN

import triton
import triton.language as tl
import torch

@triton.jit
def test_stencil(x_ptr, o_ptr):
    pid = tl.program_id(axis=0)
    rng = tl.arange(0, 4)
    x = tl.load(x_ptr + rng[:, None] + rng[None, :])
    tl.store(o_ptr + rng, tl.sum(x, axis=1))

x = torch.arange(8).cuda()
y = torch.zeros_like(x)
test_stencil[(1,)](x, y)
x, y

(tensor([0, 1, 2, 3, 4, 5, 6, 7], device='cuda:0'),
 tensor([ 6, 10, 14, 18,  0,  0,  0,  0], device='cuda:0'))

@triton.jit
def stencil_kernel_v2(
    inputs: tl.tensor,
    outputs: tl.tensor,
    shape: tl.int32,
    BLOCK_SIZE: tl.constexpr,
    NUM_OFFSETS: tl.constexpr,
    RADIUS: tl.constexpr):

    pid = tl.program_id(0)
    # Starting Offsets
    base_offs = tl.arange(0, BLOCK_SIZE) 
    inc_offs = tl.arange(0, NUM_OFFSETS) - RADIUS
    output_offsets = pid * BLOCK_SIZE + base_offs
    input_offsets = pid * BLOCK_SIZE + inc_offs[None, :] + base_offs[:, None]

    input_tensor = tl.load(inputs + input_offsets,
        mask=(input_offsets < shape and input_offsets >=0)
            and inc_offs[None, :] <= RADIUS,
        other=0.0)

    tl.store(outputs + output_offsets,
        tl.sum(input_tensor, axis=1),
        mask = output_offsets < shape)