EA: perform analysis once for post-optimization IR, and remove `IPO EA` (#51318)
Following the discussions and changes in #50805, we now consider
post-inlining IR as IPO-valid. Revisiting EA, I've realized that running
EA twice—once for computing IPO-valid escape cache and once for local
optimization analysis—is redundant. This commit streamlines the EA
process to perform the analysis just once on post-optimization IR, and
caches that result. This change also removes all interprocedural EA
code, which had significant overlap with inlining code.
---------
Co-authored-by: Julian Samaroo <jpsamaroo@jpsamaroo.me>