[mlir][affine][NFC] Count stores as int64_t in getFusionComputeCost (#243288)
When the fused slice runs a single iteration, getFusionComputeCost
records -storeCount as the cost of the insertion point's parent loop, to
subtract the stores that store-to-load forwarding is expected to remove.
storeCount was unsigned, so the negation wrapped around to nearly 2^32.
The wrapped value never reaches the result: the entry is overwritten
with the cost of the slice before the cost of the destination nest is
computed, so the stores are not subtracted either way. Count the stores
in int64_t, the type of the cost map, so that the recorded value is the
intended one. Making the subtraction take effect would change fusion
decisions and is left alone here.
Fixes static analyzer complaints about an integer overflow.
Assisted-by: Cursor