[LinalgExt] Fix attention NaN for fully-masked rows (#24178)

Attention softmax normalization produced `NaN` for fully-masked rows.

  The failing case is:
  - masked scores are all `-inf`
  - softmax numerator becomes `0`
  - softmax denominator becomes `0`
  - final normalization computes `0 / 0`

PyTorch SDPA uses `_safe_softmax`, which explicitly zeroes fully-masked
rows, so IREE should produce `0` here instead of `NaN`.

This PR handles that in both attention lowering paths:

- Standalone `iree_linalg_ext.attention` decomposition clamps the row
softmax denominator with `max(sum, 1)` before `P / sum`.
- Online attention finalization keeps the existing unmasked `(1 / sum) *
x` IR unchanged.
- Masked online attention guards the existing finalization loop so `sum
== 0` yields `0` instead of `NaN`, avoiding an extra row-level pass.

For non-fully-masked rows, the softmax denominator is unchanged: after
max subtraction, at least one term is `exp(0) = 1`, so `sum >= 1`.

  Fixes #24175.

---------

Signed-off-by: Keshav Vinayak Jha <keshavvinayakjha@gmail.com>