Explicit stores, cb lifetime ops, and side-effect-only loops

[zoecarver]

This reworks the lowering pipeline so that stores and cb lifetime ops are explicit from Python all the way down, and loops don't carry dead tensor dataflow.

• ttl.store / ttl.tile_store split: The old ttl.store is renamed to ttl.tile_store (tile-level, lives inside compute bodies). A new tensor-level ttl.store is added, emitted directly by Python's o.store(result). This means the user's store intent is in the IR from the start rather than being reconstructed later from yields.
• LowerStoreToCompute pattern: convert-ttl-to-compute now has a pattern that finds each ttl.store, locates the compute op that produces its input, and sinks to a ttl.tile_store in a ttl.compute block (this is the same logic we use to fuse eltwise ops).
• TTLInsertTileRegsSync no longer walks yields, scans for reserves, or synthesizes stores. It just inserts acquire/commit/wait/release around the existing tile_store ops. Went from ~250 lines of store inference down to straightforward insertion logic.
• Side-effect-only loops: ConvertTTLComputeToSCF generates scf.for without iter_args for outputs. No tensor.insert, no yielding updated tensors. Stores are explicit side effects via tile_store. This also simplifies ConvertTTLToTTKernel cleanup since there are no tensor.insert ops to tear down anymore.
• DST assignment refactored: TTLAssignDST now inserts all copy_tile ops at the block start (reverse arg order) via a shared createCopyTileForArg utility, instead of inserting at first-use with duplicated allocation logic.