typeintersect: fix constraintkind for non-covariant var

[N5N3]

The pre-check for constraintkind failed to predict occurs_cov (!var_occurs_invariant doesn't means occurs_cov > 0). And false constraintkind might cause circular constraints during the env contraction.

This PR fixes it by adding missing check back.
MWE added.
close #50195.

[vtjnash]

I think I misunderstood the meaning of the always_occurs_cov to mean only_cov, which was not the intent. I guess this function is actually equivalent to Core.Compiler.constrains_param && !occurs_inv

A use as a lower bound does not create a constraint on the lower bound for var:

Suggested change

	always_occurs_cov(ua->var->lb, var, 0) ||
	always_occurs_cov(ua->var->ub, var, 0) ||
	always_occurs_cov(ua->var->ub, var, 0) ||

[N5N3]

I guess they have some little difference. But the suggestions make sense (As their task looks similar).

always_occurs_cov should perform a conservative estimation on the occurs_cov after 1st intersection.
As we want to match the branch if (vb.occurs_cov && noinv) below.

var_occurs_inside is not a good template here:

1. We only perform intersect_aside on ub.
2. The union decision should not affect the conservativeness.

[maleadt]

Thanks! Confirmed that this makes the CUDA.jl hang disappear (i.e. not only the reduced MWE).