Abort local_block_diag_dot_to_dot_block_diag when n_inputs < 2

[jessegrabowski]

Closes #1864

[jessegrabowski]

is the change to pass through single inputs to block_diag a blocker for this PR?

[ricardoV94]

Nothing I'm saying is a blocker. Just trying to discuss best practices to avoid this sort of failure in the future because it's not specific to this Op or rewrite.

[jessegrabowski]

I added an eager return for the block_diag helper when n_inputs=1, and I also added a rewrite remove_useless_block_diag that kills any block diags with 1 input. That plus the n_inputs > 2 guard in the dot rewrite should put this issue to bed.

[ricardoV94]

Love it. Do you want to add the regression test nevertheless or merge as is?

[jessegrabowski]

I think no need, the regression test would be the kind of trivial test that I scold people for when their LLMs make them.