[MLIR] `affine-loop-fusion` misses producer-consumer fusion for fixed-column memref load

`affine-loop-fusion` misses a producer-consumer fusion opportunity. The producer loop initializes the full temporary buffer: `%alloc[i, j] = 1.0`. The consumer loop only reads a fixed column: `%alloc[i, 0]`. Since %alloc is locally allocated, does not escape, and is deallocated after the consumer, only the `j = 0` slice is needed.

The producer and consumer could be fused by materializing only `%alloc[i, 0]` before the load. However, `affine-loop-fusion` keeps the full producer loop nest separate from the consumer loop, resulting in a missed optimization.

### Input Program
**input.mlir**
```llvm
module {
func.func @block_removal(%arg0: memref<32xf32>) {
%cst = arith.constant 1.000000e+00 : f32
%alloc = memref.alloc() : memref<32x32xf32>
affine.for %arg1 = 0 to 32 {
affine.for %arg2 = 0 to 32 {
affine.store %cst, %alloc[%arg1, %arg2] : memref<32x32xf32>
}
}
affine.for %arg1 = 0 to 32 {
%0 = affine.load %alloc[%arg1, 0] : memref<32x32xf32>
affine.store %0, %arg0[%arg1] : memref<32xf32>
}
memref.dealloc %alloc : memref<32x32xf32>
return
}
}
```
### Command
```bash
mlir-opt input.mlir --affine-loop-fusion
```
Version: 6f54fe6b6fb22f8a335fd19b94049fc769dfea25

Start with the input.mlir reproducer and run `mlir-opt input.mlir --affine-loop-fusion` to observe that the producer remains separate from the consumer. Trace the `affine-loop-fusion` pass and its handling of fixed-column memref loads; done means the relevant producer work is fused so only the needed column is materialized.

Written by the indexing model from the issue text.