[ir] Set root block kernel after cloning

[k-ye]

I've managed to fix all the kernel == nullptr problems that I've found so far. Note that this PR alone doesn't fix all the problems. In order to run mpm99, you would still need to disable kernel fusion and constant folding. What's more interesting is that, the main constant folding workflow is fine, the problem was around

taichi/taichi/transforms/constant_fold.cpp

Lines 213 to 215 in 1402c4e

	if (kernel && kernel->program.config.debug) {
	TI_TRACE("config.debug enabled, ignoring constant fold");
	return false;

Honestly I think IRNode::get_kernel() is mostly broken in the async mode, since it depends on the assumption that root is a block..

Related issue = #1243

[Click here for the format server]

---

[codecov]

Codecov Report

Merging #1245 into master will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #1245   +/-   ##
=======================================
  Coverage   64.88%   64.88%           
=======================================
  Files          36       36           
  Lines        5041     5041           
  Branches      922      922           
=======================================
  Hits         3271     3271           
  Misses       1595     1595           
  Partials      175      175           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 7cd9a14...c229f7c. Read the comment docs.

[xumingkuan]

LGTM, but shall we modify the behavior in async mode so that the root is a Block containing an OffloadedStmt instead of a raw OffloadedStmt?

[k-ye]

but shall we modify the behavior in async mode so that the root is a Block containing an OffloadedStmt instead of a raw OffloadedStmt?

Good point. I feel like this is a bit ad-hoc, but I cannot think of how this could break things in any other way.. Allow me to merge this one first