ttnn.sampling support for Llama3 8b N150

[ipotkonjak-tt]

Describe the bug
ttnn.sampling invoked with tensor of the shape [1, 1, 32, 128256] (tensor size of the llama3 8b output on N150) isn't supported.

When called the following error occurs:
Always | FATAL | TODO: add support for multi-paged buffer with page size > 64KB

To Reproduce

@pytest.mark.parametrize(
    "shape",
    [
        # [1, 1, 32, 32 * 8], # PASS
        [1, 1, 32, 128256], # llama3 8b n150 - FAIL
    ],
)
@pytest.mark.parametrize("k", [[1] * 32])  # Example of per-user k
@pytest.mark.parametrize("p", [[1.0] * 32])  # Example of per-user p
@pytest.mark.parametrize("seed", [2024])
@pytest.mark.parametrize(
    "sub_core_grids", [ttnn.CoreRangeSet({ttnn.CoreRange(ttnn.CoreCoord(0, 0), ttnn.CoreCoord(8 - 1, 4 - 1))})]
)
def test_argmax(shape, k, p, seed, device, sub_core_grids, use_program_cache):
    device.enable_async(True)
    torch.manual_seed(seed)

    # Input tensors
    input_values = torch.randn(shape)
    input_indices = torch.arange(0, shape[-1], dtype=torch.int32).expand(shape)

    # TTNN input tensors
    input_values_tensor = ttnn.from_torch(input_values, device=device, dtype=ttnn.bfloat16, layout=ttnn.TILE_LAYOUT)
    input_indices_tensor = ttnn.from_torch(input_indices, device=device, dtype=ttnn.int32, layout=ttnn.ROW_MAJOR_LAYOUT)

    # Argmax compute
    argmax_output = torch.argmax(input_values, -1, keepdim=True).transpose(-1, -2)

    # ttnn.sampling as argmax
    ttnn_output = ttnn.sampling(
        input_values_tensor, input_indices_tensor, k=k, p=p, seed=seed, sub_core_grids=sub_core_grids
    )
    output = ttnn.to_torch(ttnn_output)

    assert_with_pcc(output, argmax_output, 0.9999)

Additional context
This feature is needed to avoid host fallback and make data parallel execution of llama models performant.