Block::kernel isn't set correctly in the async computation

[k-ye]

Describe the bug

Block::kernel isn't set correctly in the async computation. In the default mode, this works because the root block maps to a Taichi kernel, which has its kernel set. In the async mode, however, the root block maps to an offloaded task.

Other problems so far:

• fuse need to respect memory barrier implicitly established by the kernel boundary.
• constant folding??

[k-ye]

One thing I forgot to mention: We should probably be a bit more careful when using default args, e.g.

taichi/taichi/ir/transforms.h

Line 37 in 24f6836

void lower_access(IRNode *root, bool lower_atomic, Kernel *kernel = nullptr);

I feel like this has somehow complicated the code flow when debugging these things.

[xumingkuan]

I wonder why we still need to pass these parameters given that we can get the Kernel * from an IRNode *...

[k-ye]

I wonder why we still need to pass these parameters given that we can get the Kernel * from an IRNode *...

Well, this is probably added before we added IRNode::get_kernel()? But given that it's not working reliably...

[k-ye]

This particular bug is fixed by #1245..