[ORC] WaitingOnGraph perf: faster dependence propagation. (#183272)

This commit replaces the core dependence propagation algorithm in
WaitingOnGraph to avoid worst-case behavior in the common case where
dependence graphs are sparse. This algorithm showed up as the underlying
cause of the bug in https://github.com/llvm/llvm-project/issues/179611.

For each call to MaterializationResponsibility::notifyEmitted,
WaitingOnGraph would build the transitive closure of all SuperNodes
whose "waiting on" relationships were affected by the newly emitted
symbols, then propagate any remaining unemitted dependencies through
this transitive closure graph. This approach is simple, but pushes the
algorithm towards n^2 complexity even for sparse dependence graphs.

The new propagation algorithm:
1. Inverts the edge direction in the SymbolDependenceMap data structure:
SymbolDepMap[SN] now contains the set of SuperNodes that depend on SN,
rather than the set that SN depends upon.

2. Pushes dependencies through the SymbolDepMap iteratively until it
reaches a fixed point.

This updated algorithm converges much more quickly than the original for
the testcase reported in the issue, and for other cases tested so far.

No testcase yet, as this only affects performance. I've filed a
follow-up issue, https://github.com/llvm/llvm-project/issues/183251, to
track construction of performance testing infrastructure for
WaitingOnGraph.

Should fix https://github.com/llvm/llvm-project/issues/179611