Add replicated tensor type and handle conv2d with sharded input channels

[stellaraccident]

Discussed offline about possibly changing __eq__ to something like is_deep_equivalent and documenting specifically what that means.

[sogartar]

@stellaraccident I renamed __eq__ to is_deep_equal.

[sogartar]

I rebased and resolved conflicts.

[sogartar]

@stellaraccident, I ran into this problem that when loading the dataset if we mem map, then even when deleting the Python objects, the underlying files are still open. I see that the files are memory mapped in IREE. I assume we are missing somewhere an ownership relation.