CCL operations with replica_groups=1

[tapspatel]

When running through compiler passes in tt-forge, if the CCL operation is done via the following configuration

1x8

all_gather on cluster_axis=0

Therefore, we are doing all_gather on each device with itself. This doesn't produce anything useful computationally but for generality, it would be nice to have an optimization that could return back the original tensors unchanged. This is very useful for being able to run a sweep of tests with various different configurations that build on-top of tt-metal. Currently, it fails with the following error

Error, All-gather was unable to identify either a sender or receiver device ID and atleast one must be identified for a valid all-gather configuration. The input mesh tensor or all-gather arguments may be incorrect

Sample test case:

import ttnn
import torch

mesh_device = ttnn.open_mesh_device(ttnn.MeshShape(1, 8))

torch_input_tensor = torch.rand((1, 1, 8192,800), dtype=torch.bfloat16)

mesh_tensor = ttnn.from_torch(
    torch_input_tensor,
    layout=ttnn.TILE_LAYOUT,
    device=mesh_device,
    mesh_mapper=ttnn.ShardTensor2dMesh(mesh_device, mesh_shape=list(mesh_device.shape), dims=(2, 3)),
)

output_tensor = ttnn.all_gather(
    mesh_tensor,
    dim=2,
    cluster_axis=0,
    mesh_device=mesh_device,
    topology=ttnn.Topology.Linear,
)

ttnn.close_mesh_device(mesh_device)