Make cuTile intrinsics 1-based

[maleadt]

This is an experiment.

Summary: Move 1-based to 0-based index lowering from the DSL (operations.jl) into a compiler rewrite pass (index_lower_pass!). The load/store/gather operations now pass through Julia's natural 1-based indices, and the pass inserts subi(elem, 1) for each index element in load_partition_view and store_partition_view calls during compilation.

The motivation here is to simplify the IR we emit (the layernorm example currently emits 4x as much SASS instructions as cuTile Python does). A large part of that IR cruft comes from the repetitive +1/-1 we do as part of the 0-based to 1-based index conversion. For example, bid(1) returns a 1-based index (via addi(blockId_x, 1)), then each load/store call emits its own subi(..., 1) to undo it, resulting in 3 redundant constant+subi pairs:

  %1 = addi %blockId_x, %cst_1_i32          // bid(1) = blockId_x + 1
  %2 = subi %1, %cst_1_i32_7                // load a: undo +1
  load_view_tko ... %pview[%2]
  %3 = subi %1, %cst_1_i32_9                // load b: undo +1 again
  load_view_tko ... %pview_8[%3]
  %5 = subi %1, %cst_1_i32_12               // store c: undo +1 again
  store_view_tko ... %pview_13[%5]

On this branch, the indices passed around are kept 1-based, and a pass converts them late at the load/store boundary, result in significantly simpler IR:

  %1 = addi %blockId_x, %cst_1_i32          // bid(1) = blockId_x + 1
  load_view_tko ... %pview[%1]              // pass lowered index, addi+subi cancelled
  load_view_tko ... %pview_7[%1]
  store_view_tko ... %pview_10[%1]

Although this is nice, it both doesn't improve performance, and makes it harder to compare our IR to cuTile Python's. So not sure we want this.

[maleadt]

Algebraic simplifications have turned out powerful enough to not need this.