[inductor] Generalize pointless_cumsum_replacement pattern (#108373)
The current pattern transforms:
```
ones([x, y]).cumsum(1) -> arange(1, 1 + y).expand([x, y])
```
but this generalizes it to
```
full(shape, fill_value).cumsum(d) ->
(fill_value * arange(1, 1 + shape[d])).view([1..., shape[d], 1...]).expand(shape)
```
So we handle any fill value, any number of dimensions, and broadcasting to any dimension.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108373
Approved by: https://github.com/lezcano
