Alias-aware token threading for better parallelism

[maleadt]

All memory operations (loads, stores, atomics) are threaded through a single global token chain. This is correct but conservative—operations on independent arrays are serialized unnecessarily.

Proposed improvement

Implement alias-aware token threading:

1. Alias analysis: Compute which pointers may refer to the same memory region (alias sets)
2. Per-set token chains: Thread tokens only between operations that may alias
3. Loop parallel stores: Identify stores in loops with non-overlapping indices across iterations—these can skip token dependencies entirely

Why

The current sequential approach prevents parallelism between independent memory operations. For example, loading from array a and storing to array b don't need ordering constraints if they're provably disjoint. Alias-aware threading preserves correctness while enabling the hardware to execute independent operations concurrently.

Reference implementation

cuTile Python implements this in:

• https://github.com/NVIDIA/cutile-python/blob/main/src/cuda/tile/_passes/alias_analysis.py — dataflow analysis propagating alias sets until fixed-point
• https://github.com/NVIDIA/cutile-python/blob/main/src/cuda/tile/_passes/token_order.py — maps tokens to alias sets via TokenKey, with special handling for loop-parallel stores