Use [:mintap] instead of [:-n_tap_done], to make it use the same was as other place to help subtensor opt.

an existing tests.

Also add assert when test_value are available.

This in combination with PushOutSeqScan make shape error! as PushOutSeqScan make the assumption that was inforced by the scan() fct that all input sequence have the same length.

TODO: make a test for this.

The return value isn't used in this case, but consistency is good.

That allow to know easily witch tag to use to disable an opt.

This make the printing of the reason be simpler/more readable.

This allow to easily know how to disable that specific optimization.

[WIP] Adding GpuJoin.c_code to cuda/basic_ops.py

doc: fix typo

Doc

Crash gpu reduce 1111