Add loop parallel store optimization and DCE pass

[maleadt]

Implements two compiler passes that clean up unnecessary token overhead from the alias-aware token ordering pass (#89),
matching cuTile Python's output:

1. Loop parallel store optimization: Stores in for-loops with injective indices (using the induction variable) use the parent scope's token instead of a loop-carried token, breaking the token dependency chain through
   the loop. Matches Python's _try_loop_parallel_store.
2. Dead code elimination: General-purpose DCE using dependency graph reachability analysis. Removes dead token carries, join_tokens, and unused instructions left behind by the parallel store optimization. Uses Julia's efunc effect annotations to classify intrinsic side effects. Matches Python's dead_code_elimination_pass.

Together these eliminate all dead token loop carries and join_tokens from memory-bound kernels like layernorm, producing token IR structurally identical to cuTile Python (all ops use the root token, zero loop-carried tokens).

Closes #146

[maleadt]

Not perfect yet, but I have a couple of things building on top of this so let's merge already.