Add rfactor patterns for index-first argmax/argmin

[alexreinking]

Add patterns to the associative ops table to allow writing argmin/argmax as (index, value) pairs. Previously, rfactor supported only (value, index) because it would fail to prove associativity.

Doing this required reworking the code that converted a matched expression back into a pattern over the new variable names. It did this completely backwards for some reason: trying to match large fragments of IR and replacing them with the placeholder names. But we just matched the pattern in the table! So I'm proceeding by renaming the table patterns directly. I don't see why this shouldn't work, but I've only run correctness_rfactor and _test_internal locally, so we'll see what the bots say.

Added tests for the inline reduction helpers, too.

ATTN @stevenraphael

[alexreinking]

"Fun" fact: without this patch, we get a segfault when HL_DEBUG_CODEGEN is set to print this message. When the associative op is in the second tuple component, only y1 is in the result, yet the lookup itself will add an entry to for y0, with an undefined Expr for the value. Later, the expr_uses_vars check will try to visit the undefined Expr and a null pointer gets dereferenced.